<?xml version="1.0" encoding="UTF-8"?><rss version="2.0" xmlns:content="http://purl.org/rss/1.0/modules/content/">
  <channel>
    <title>表图</title>
    <link>https://wechat2rss.xlab.app/feed/657873c2f534ea1c50875c8657bc405270ce7cd0.xml</link>
    <description>坚信少却更好，坚定元认知传播，坚持安全美学，拿个主题，讲300秒就够了。&#xA;(wechat feed made by @ttttmr https://wechat2rss.xlab.app)</description>
    <managingEditor> (表图)</managingEditor>
    <image>
      <url>https://wx.qlogo.cn/mmhead/Q3auHgzwzM4bPknPOgOqIB6IcWHuyHjhSyorB6EsibPchqfiaY8cCfcQ/0</url>
      <title>表图</title>
      <link>https://wechat2rss.xlab.app/feed/657873c2f534ea1c50875c8657bc405270ce7cd0.xml</link>
    </image>
    <item>
      <title>[译苑雅集 Vol. 19] 把界面交给 AI，Salesforce 反而更难被替代？</title>
      <link>https://mp.weixin.qq.com/s?__biz=MzUzOTI4NDQ3NA==&amp;mid=2247485197&amp;idx=1&amp;sn=17776eea48e067f5c3ce68b4fd591f34</link>
      <description>Salesforce 把界面交给 Claude，保留数据、业务规则和执行能力。“无头 SaaS”可能加深客户依赖、改善 CRM 数据，也让双方的合作埋下未来平台竞争的伏笔。</description>
      <content:encoded><![CDATA[<p><span>四楼南侧东</span> <span>2026-09-02 23:14</span> <span style="display: inline-block;">北京</span></p>






  
  <p><img src="https://wechat2rss.xlab.app/img-proxy/?k=79f4c933&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fmmbiz_jpg%2FNBNzfW7CHBXbVNcVHIqwvdftbwuZAic7MUvvpUaTPjQwqB7lh7Oo7mTPtQNzlvGQomal4988lNCAujP4TZ5A3dOQrT14ZupwWjK9SicpEfpLc%2F0%3Fwx_fmt%3Djpeg"/></p>
  <p>Salesforce 把界面交给 Claude，保留数据、业务规则和执行能力。“无头 SaaS”可能加深客户依赖、改善 CRM 数据，也让双方的合作埋下未来平台竞争的伏笔。</p>
  <p style="text-align: center;" nodeleaf=""><img data-aistatus="1" class="rich_pages wxw-img" data-ratio="0.562799043062201" data-s="300,640" data-type="png" data-w="1672" type="block" data-imgfileid="100001539" src="https://wechat2rss.xlab.app/img-proxy/?k=7c068746&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fmmbiz_png%2FNBNzfW7CHBWkETI2IMgsGeJos8DHKicibCO4hNymgKEvIFrTnuxlu6aU3AOb6coFH8JQSFVYpibOC0pUWnvicWWnetiaFzwtWtqr1wze8rrYDAPI%2F640%3Fwx_fmt%3Dpng%26from%3Dappmsg"/></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);" data-pm-slice="0 0 []"><span leaf="">作者：Saanya Ojha</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">时间：2026 年 08 月 29 日</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">原文：</span><span leaf=""><a href="https://saanyaojha.substack.com/p/the-claudeforce-awakens" target="_blank">https://saanyaojha.substack.com/p/the-claudeforce-awakens</a></span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);" data-pm-slice="0 0 []"><span leaf="">SaaS 回来了，而且势头很猛。今年大部分时间里，软件股都备受市场冷落——还记得那场“SaaS 末日”（SaaSpocalypse）吗？如今，市场对软件股的信心突然又回来了。过去五天，ServiceNow 上涨约 12%，Salesforce 上涨约 25%。SaaS 行业的估值倍数也普遍明显回升，已远高于仅仅三个月前触及的低点。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">本周这轮反弹的催化剂之一，是 Salesforce 一个在战略上颇值得玩味的动作：它把自己搬进了 Claude。这是 Salesforce 今年早些时候推出的“无头”（headless）战略的自然延伸。基本思路很简单，却相当激进：将保存企业权威业务记录的系统（system of record）与用户界面拆开。用马克·贝尼奥夫（Marc Benioff）那句更让人印象深刻的话来说：“用户界面就是 AI。”</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这项把 Salesforce 带进 Claude 的合作名为 Claudeforce，首款插件 Salesforce in Claude 内置了 37 项销售技能（skills），涵盖销售管线复盘、会前准备、客户分析和商机健康度评估。Claude 可以结合 Salesforce 的实时数据、Slack 信息以及其他已接入的上下文进行推理，然后真正动手：更新商机、创建营销活动、执行工作流。底层的权限控制、业务逻辑和治理规则，仍由 Salesforce 负责执行。</span></p><figure style="width: 728px;margin: 0px auto;"><p><picture style=";"><span leaf=""><img class="rich_pages wxw-img" data-aistatus="1" data-imgfileid="100001548" data-ratio="1.275" width="621" data-type="jpeg" data-w="960" height="791.775" sizes="100vw" style="display: block;max-width: 100%;height: auto;margin: 0px auto;" src="https://wechat2rss.xlab.app/img-proxy/?k=cd5309fc&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fmmbiz_jpg%2FNBNzfW7CHBVRh27K4X41uTDEht2hd6tNNIyQuZzfiaQevYWnR9z8Rh6p7KvWcLC0Fk2mwzQssC4mQrI1EmhmjnticbOh1VNtFMrJhLTiauSpB0%2F640%3Fwx_fmt%3Djpeg%26from%3Dappmsg"/></span></picture></p></figure><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">支撑这一切的是 Salesforce 新推出的 Headless 360 / AIforce 架构：通过 MCP、API 和其他编程接口，向智能体开放数据、元数据、工作流和操作。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">Salesforce 试图让 MCP 成为自己的下一个“API 时刻”。当年，Salesforce 能建立起强大的平台，是因为第三方开发者可以通过 API 在其平台上构建应用。Headless 360 希望在智能体时代重现这一成功：让智能体能够通过 MCP 发现 Salesforce 中的每个对象、工作流和操作，同时在底层保留 Salesforce 的治理机制。Salesforce 也明确把 Headless 360 比作自己 27 年前率先推出的 Web API。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">此前，Salesforce 的商业模式是：</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><strong style=";"><span leaf="">掌握数据 → 掌握应用 → 掌握用户界面 → 按席位收费。</span></strong></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">现在，它正在转向：</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><strong style=";"><span leaf="">掌握数据、语义、权限和操作 → 让别人掌握用户界面 → 从完成的工作中获得收入。</span></strong></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这意味着 Salesforce 将成为一家很不一样的公司。至少在本周，华尔街很买账。周四，Salesforce 股价上涨 22.6%，创下六年来最大单日涨幅。</span></p><figure style="width: 728px;margin: 0px auto;"><p><picture style=";"><span leaf=""><img class="rich_pages wxw-img" data-aistatus="1" data-imgfileid="100001546" data-ratio="0.902970297029703" width="597" data-type="jpeg" data-w="1010" height="539.0732673267327" sizes="100vw" style="display: block;max-width: 100%;height: auto;margin: 0px auto;" src="https://wechat2rss.xlab.app/img-proxy/?k=73056066&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fsz_mmbiz_jpg%2FNBNzfW7CHBX6phqFlCnHSJ4W5SYtHA5km6Ppr3gm5dzGsKwV2wZF3jucMYTLpPSRBlNROUYzbYOBuFjjORiaSBuUJVghlql5HV2XxX2mhNMU%2F640%3Fwx_fmt%3Djpeg%26from%3Dappmsg"/></span></picture></p><p><picture style=";"></picture></p><p><picture style=";"><span leaf="">当然，这轮上涨也不能全归功于 Claudeforce。Salesforce 同时交出了一份强劲的季度业绩：当期剩余履约义务（cRPO）同比增长 14%，净新增年度订单额的增速创四年来新高；Agentforce 与 Data 360 的年度经常性收入（ARR）合计接近 39 亿美元，其中 Agentforce 的 ARR 超过 15 亿美元。管理层还上调了 2027 财年的收入指引。</span></picture></p></figure><figure style="width: 728px;margin: 0px auto;"><p><picture style=";"><span leaf=""><img class="rich_pages wxw-img" data-aistatus="1" data-imgfileid="100001547" data-ratio="1.1788079470198676" width="604" data-type="jpeg" data-w="604" height="712" sizes="100vw" style="display: block;max-width: 100%;height: auto;margin: 0px auto;" src="https://wechat2rss.xlab.app/img-proxy/?k=7a9ecde7&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fsz_mmbiz_jpg%2FNBNzfW7CHBUqJFt64agKicHWmyvibxqspFbZT5Tq3RoHb4qV4icrw50qsibuqdBel5zc32mSDicn8sId0tYjSSCoXibibfyZygM5aIsDvA4l9qROj0%2F640%3Fwx_fmt%3Djpeg%26from%3Dappmsg"/></span></picture></p><p><picture style=";"></picture></p><p><picture style=";"><span leaf="">此外，账面收益也让财报增色不少：Salesforce 对 Anthropic 的投资产生了约 27 亿美元未实现收益；包括这笔收益在内的战略投资净收益，为 5.90 美元的调整后每股收益（EPS）贡献了 2.53 美元。但这轮上涨的幅度，似乎超出了单纯业绩超预期所能解释的范围。它更像是市场在为 Salesforce 的新叙事重新定价。</span></picture></p></figure><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">对此，我有几点想法。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;overflow-wrap: anywhere;word-break: keep-all;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">1. SaaS 可以放弃用户界面，却离不开业务记录系统</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">“SaaS 末日论”的逻辑可以简化为：智能体能够操作软件，所以智能体会取代软件。但一个能够推理的 AI，与一个清楚以下问题的企业系统，仍有很大差别：</span></h1><ul style="margin-top: 0px;padding: 0px;" class="list-paddingleft-1"><li style="margin: 8px 0px 0px 32px;"><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);box-sizing: border-box;padding-left: 4px;"><span leaf="">Acme 公司是哪家客户；</span></p></li><li style="margin: 8px 0px 0px 32px;"><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);box-sizing: border-box;padding-left: 4px;"><span leaf="">它买过什么；</span></p></li><li style="margin: 8px 0px 0px 32px;"><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);box-sizing: border-box;padding-left: 4px;"><span leaf="">谁有权修改它的合同；</span></p></li><li style="margin: 8px 0px 0px 32px;"><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);box-sizing: border-box;padding-left: 4px;"><span leaf="">你们公司如何定义销售管线；</span></p></li><li style="margin: 8px 0px 0px 32px;"><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);box-sizing: border-box;padding-left: 4px;"><span leaf="">哪种折扣需要副总裁批准；</span></p></li><li style="margin: 8px 0px 0px 32px;"><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);box-sizing: border-box;padding-left: 4px;"><span leaf="">当有人点击“成交”（Close Won）时，系统究竟会执行哪些操作。</span></p></li></ul><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">Salesforce 提供的是企业业务状态、业务语义、授权规则和执行能力。在智能体时代，这些可能构成一道比用户界面更重要的护城河。AI 或许会让挡风玻璃变成随处可买的标准件，同时让发动机更值钱。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;overflow-wrap: anywhere;word-break: keep-all;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">2. Salesforce 也在大幅提升 Anthropic 的地位</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这项合作中，藏着一场颇有意思的权力博弈。Salesforce 希望成为支撑所有模型的企业级运行框架（harness），Claude 则希望成为所有企业应用之上的界面。这两个位置今天相互补充，明天却可能相互竞争。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">如果员工每天早上都从 Claude 开始工作，无须先打开 Salesforce、Workday、ServiceNow、Snowflake、Jira，或者十来个浏览器标签页，那么 Claude 就成了汇聚企业需求的新入口。这会是一个极有价值的位置。Salesforce 赌的是：只要保住底层权威的数据模型和事务处理层，失去用户界面也可以接受。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">Anthropic 想占据员工的工作入口，Salesforce 想占据企业的业务核心。Claudeforce 让双方握手合作，也可能成为它们未来的战场。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;overflow-wrap: anywhere;word-break: keep-all;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">3. “无头 SaaS”反而可能提高切换成本</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">乍看之下，去掉用户界面，似乎会让产品更趋同质化。但不妨设想：企业内部成百上千个智能体持续发现并调用 Salesforce 的能力，而机器生成的各类工作流中，到处嵌着 Salesforce 的对象、元数据和业务规则。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">如今，切换 Salesforce 的成本来自定制开发、系统集成和对现有报表的依赖，光是让两万名员工重新接受培训，就足以让人心力交瘁。未来，切换成本可能更多来自解除数百万条机器之间对 Salesforce 业务语义的依赖。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这些依赖关系更难被看见，也可能更难解除。耐人寻味的是，人们通过 Claude 访问 Salesforce，可能让 Salesforce 在企业中嵌得更深，同时又越来越不显眼。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;overflow-wrap: anywhere;word-break: keep-all;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">4. 我们可能正在重现浏览器与应用的关系</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">网页浏览器是个有意思的历史参照。浏览器后来成为访问各类网站的通用界面。Claude 可能在企业里扮演类似角色：Salesforce、Workday 和 ServiceNow 依然存在，只是由 Claude 根据需要呈现相关信息，并调用它们的功能。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">如果如此，问题就变成了：谁能占据 Google 的位置，谁又会落在 HTTP 服务器的位置？Salesforce 的商业体量可能依然庞大，而 Anthropic 则掌握信息发现、用户意图和用户注意力。这也正是今天的合作中已经埋下明天平台冲突的原因。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;overflow-wrap: anywhere;word-break: keep-all;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">5. 生成式 UI 让应用成为按需呈现的能力</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">Salesforce 描述过这样一种场景：Claude 根据每位销售人员负责的客户、销售管线和企业的实时上下文，生成定制化的仪表盘。这比对话式界面更激进。过去，Salesforce 的产品经理决定有哪些页面，再由工程师把它们做出来。如今，应用实际上可以在用户发起查询的那一刻即时“编译”出来。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这意味着，SaaS 可能从“软件厂商为不同用户角色设计工作流”，转向“模型为眼前的具体情境生成界面”。如果如此，构建设计系统、仪表盘、导航树和报表界面所需的开发工作量，都可能大幅减少。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;overflow-wrap: anywhere;word-break: keep-all;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">6. 智能体或许终于能解决 CRM 的原罪</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="" style=";">客户关系管理（CRM）系统一直有一个结构性缺陷：数据库的质量取决于人们有多愿意更新它。销售人员更愿意去做销售，而非替销售副总裁维护一个结构规整的数据库，这完全合情合理。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span style=";"><span leaf="">智能体改变了这一点。它们可以持续观察会议、邮件、Slack 和文档中的信息，将其中反映的业务活动与 Salesforce 的记录核对，并据此更新数据。由此会形成一个颇有意思的飞轮：</span></span><strong style=";"><span leaf="">更好的智能体 → 更完整的 CRM 数据 → 更好的上下文 → 更好的智能体。</span></strong><span style=";"><span leaf=""> 用户花在手动操作 Salesforce 上的时间更少了，AI 却可能因此让底层数据变得更好。</span></span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">在 SaaS 的第一个时代，我们为每一种业务流程都加上了图形界面。到了下一个时代，我们可能又会逐步卸下这些界面，让应用成为支撑智能的无头基础设施。</span></p><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);"><span leaf="">Claudeforce 最耐人寻味之处在于：Salesforce 也许恰恰要让自己越来越隐形，才能维持自己的重要地位。</span></p><p style="display: none;"><mp-style-type data-value="3"></mp-style-type></p>



<p><a href="https://wechat2rss.xlab.app/link-proxy/?k=12ef04e2&amp;r=1&amp;u=https%3A%2F%2Fmp.weixin.qq.com%2Fs%3F__biz%3DMzUzOTI4NDQ3NA%3D%3D%26mid%3D2247485197%26idx%3D1%26sn%3D17776eea48e067f5c3ce68b4fd591f34">跳转微信打开</a></p>
]]></content:encoded>
      <pubDate>Wed, 02 Sep 2026 23:14:00 +0800</pubDate>
    </item>
    <item>
      <title>OpenAI - Hugging Face 事件最终技术报告：当上千个 Agent 开始互相授权，谁还能让它们停下？</title>
      <link>https://mp.weixin.qq.com/s?__biz=MzUzOTI4NDQ3NA==&amp;mid=2247485185&amp;idx=1&amp;sn=0a952f0343c959bc6785c2330ec25025</link>
      <description>上千个 Agent 在数日内形成分工、规则和授权关系。OpenAI - Hugging Face 事件最终技术报告揭示的深层风险，是命令权向同伴网络漂移，而人类没有及时行使停止权。</description>
      <content:encoded><![CDATA[<p><span>四楼南侧东</span> <span>2026-09-01 17:43</span> <span style="display: inline-block;">北京</span></p>






  
  <p><img src="https://wechat2rss.xlab.app/img-proxy/?k=f0f2ecb8&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fsz_mmbiz_jpg%2FNBNzfW7CHBXGcMlmCNWa2SYqQoWWBgQ1QEiaTbDXPeb7wAneEQicZk4u1iaxZvXzXxFictddM6hX1twFbmgRxxQdUMgw4d7LfI8bBxhKJZGUibgE%2F0%3Fwx_fmt%3Djpeg"/></p>
  <p>上千个 Agent 在数日内形成分工、规则和授权关系。OpenAI - Hugging Face 事件最终技术报告揭示的深层风险，是命令权向同伴网络漂移，而人类没有及时行使停止权。</p>
  <p style="text-align: center;" nodeleaf=""><img data-aistatus="1" class="rich_pages wxw-img" data-ratio="0.562799043062201" data-s="300,640" data-type="png" data-w="1672" type="block" data-imgfileid="100001533" src="https://wechat2rss.xlab.app/img-proxy/?k=455f417f&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fsz_mmbiz_png%2FNBNzfW7CHBXVnsgJ3pOzNjJRGJzIQTNI1poAK8ic7zOPDnGv6EGceWkvUyIWhuZauQlDn44sDaFAkIsuLdTia1QicDo6aZEf3hvfwoDUyibtyoI%2F640%3Fwx_fmt%3Dpng%26from%3Dappmsg"/></p><p><span style="color: rgb(54, 55, 55);font-family: Spectral, serif, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-size: 19px;font-style: normal;font-variant-ligatures: normal;font-variant-caps: normal;font-weight: 400;letter-spacing: normal;orphans: 2;text-align: start;text-indent: 0px;text-transform: none;widows: 2;word-spacing: 0px;-webkit-text-stroke-width: 0px;background-color: rgb(255, 255, 255);text-decoration-thickness: initial;text-decoration-style: initial;text-decoration-color: initial;display: inline !important;float: none;" data-pm-slice="0 0 []"><span leaf="">过去一个多月，我围绕 OpenAI 的 AI Agent 入侵 Hugging Face 写了三篇文章，分别讨论了它们的端到端攻击能力、共享留言板形成的“攻击组织”，以及奖励黑客、Harness 与超级对齐。</span></span></p><p><span style="color: rgb(54, 55, 55);font-family: Spectral, serif, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-size: 19px;font-style: normal;font-variant-ligatures: normal;font-variant-caps: normal;font-weight: 400;letter-spacing: normal;orphans: 2;text-align: start;text-indent: 0px;text-transform: none;widows: 2;word-spacing: 0px;-webkit-text-stroke-width: 0px;background-color: rgb(255, 255, 255);text-decoration-thickness: initial;text-decoration-style: initial;text-decoration-color: initial;display: inline !important;float: none;" data-pm-slice="0 0 []"><span leaf=""><a class="normal_text_link mp_article_text_link" target="_blank" style="" href="https://mp.weixin.qq.com/s?__biz=MzUzOTI4NDQ3NA==&amp;mid=2247485126&amp;idx=1&amp;sn=99ef9715d3bc202f8e635b148ce7591b&amp;scene=21#wechat_redirect" textvalue="[译苑雅集Vol. 17] 为了在测试中作弊，OpenAI 的最强 AI 自主入侵了 Hugging Face" data-itemshowtype="0" linktype="text" data-linktype="2">[译苑雅集Vol. 17] 为了在测试中作弊，OpenAI 的最强 AI 自主入侵了 Hugging Face</a></span></span></p><p><span style="color: rgb(54, 55, 55);font-family: Spectral, serif, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-size: 19px;font-style: normal;font-variant-ligatures: normal;font-variant-caps: normal;font-weight: 400;letter-spacing: normal;orphans: 2;text-align: start;text-indent: 0px;text-transform: none;widows: 2;word-spacing: 0px;-webkit-text-stroke-width: 0px;background-color: rgb(255, 255, 255);text-decoration-thickness: initial;text-decoration-style: initial;text-decoration-color: initial;display: inline !important;float: none;" data-pm-slice="0 0 []"><span leaf=""><a class="normal_text_link mp_article_text_link" target="_blank" style="" href="https://mp.weixin.qq.com/s?__biz=MzUzOTI4NDQ3NA==&amp;mid=2247485150&amp;idx=1&amp;sn=1a8cdb68ef696a30ab24325e1101b852&amp;scene=21#wechat_redirect" textvalue="OpenAI 披露更多 Hugging Face 入侵事件信息：一群 Agent 自发形成了“攻击团伙”" data-itemshowtype="0" linktype="text" data-linktype="2">OpenAI 披露更多 Hugging Face 入侵事件信息：一群 Agent 自发形成了“攻击团伙”</a></span></span></p><p><span style="color: rgb(54, 55, 55);font-family: Spectral, serif, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-size: 19px;font-style: normal;font-variant-ligatures: normal;font-variant-caps: normal;font-weight: 400;letter-spacing: normal;orphans: 2;text-align: start;text-indent: 0px;text-transform: none;widows: 2;word-spacing: 0px;-webkit-text-stroke-width: 0px;background-color: rgb(255, 255, 255);text-decoration-thickness: initial;text-decoration-style: initial;text-decoration-color: initial;display: inline !important;float: none;" data-pm-slice="0 0 []"><span leaf=""><a class="normal_text_link mp_article_text_link" target="_blank" style="" href="https://mp.weixin.qq.com/s?__biz=MzUzOTI4NDQ3NA==&amp;mid=2247485161&amp;idx=1&amp;sn=1c3e5f54f27433991d8930d6c97b8b97&amp;scene=21#wechat_redirect" textvalue="从奖励黑客到超级对齐：AI 为何会完成任务，却背离人类意图" data-itemshowtype="0" linktype="text" data-linktype="2">从奖励黑客到超级对齐：AI 为何会完成任务，却背离人类意图</a></span></span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);" data-pm-slice="0 0 []"><span leaf="">前两天，OpenAI 和 METR / Redwood Research 分别发布了 38 页的最终技术报告和 91 页的独立调查，把技术时间线、Agent 轨迹和内部响应过程拼得更完整。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span style=";"><span leaf="">OpenAI: OpenAI – Hugging Face Incident Technical Report</span></span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span style=";"><span leaf=""><a href="https://cdn.openai.com/pdf/67869394-cb91-4c12-888c-5cbd85c7814c/OpenAI-Hugging-Face%20Incident-Technical-Report.pdf" target="_blank">https://cdn.openai.com/pdf/67869394-cb91-4c12-888c-5cbd85c7814c/OpenAI-Hugging-Face%20Incident-Technical-Report.pdf</a></span></span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span style=";"><span leaf="">METR &amp; Redwood Research: Brief independent investigation of agents’behavior , reasoning and collaboration in the OpenAI / Hugging Face hacking incident</span></span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span style=";"><span leaf=""><a href="https://metr.org/hugging-face-incident-report-aug-2026.pdf" target="_blank">https://metr.org/hugging-face-incident-report-aug-2026.pdf</a></span></span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">从 5 月开始出现的留言板、违规联网和基础设施探测，可以看作后来事件的行为前兆。不过，OpenAI 明确认定，5 月另有一批访问 Hugging Face 的 Agent 工作负载与 7 月入侵无关。承载这次入侵协作的主要留言板，是在旧环境重建后于 7 月 8 日形成的，到 7 月 13 日攻击基本结束。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">关于攻击动机，此前最直观的解释是，一群参加网络安全评测的 Agent 为了通过测试，跑到 Hugging Face 偷答案。最终报告显示，它们很早就逆向算出了评测所需的 flag，根本不需要入侵 Hugging Face 才能交卷。推动攻击不断升级的，是 Agent 对评测机制的错误想象：它们相信存在一个会审查完整解题过程的“因果评分器”，而这个评分器实际上从未部署。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span style=";"><span leaf="">这些新证据指向了一种值得警惕的现象：</span></span><strong style=";"><span leaf="">授权漂移</span></strong><span style=";"><span leaf="">。它指的是，人类给定的目标和权限在运行过程中逐渐失去最高约束力，同伴任务、留言板规则、协调者声望和签名消息开始取得事实上的命令权。部分 Agent 明知行动已经越界，仍然接受了来自同伴网络的新任务。</span></span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这起事件呈现出两种互为镜像的授权失真。Agent 一侧产生了本不该拥有的横向授权；人类一侧已经看到违规联网、端口扫描、异常通信和内部基础设施失陷，却没有及时形成跨越安全、对齐和研究团队的事故判断，也没有行使覆盖整个系统的停止权。最终报告留下的最重要问题，正是控制权如何在两套组织之间发生了偏移。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;overflow-wrap: anywhere;word-break: keep-all;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">一个不存在的评分器，把作弊推向了更危险的路径</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">OpenAI 当时正在运行 ExploitGym，一套测试模型网络攻击能力的内部评测。Agent 会被放进隔离环境，获得一个指定靶标和一项攻击任务，成功取得 flag 就算完成。很多任务事实上无法按预期完成：898 项任务中有 198 项从未被任何模型解出，ExploitGym 的设计者则估计约有 30% 至 40% 的任务不可解。进入留言板讨论的任务中，93% 来自从未成功完成的那一组。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">长期卡在这些任务上，让 Agent 开始研究评测本身。主要留言板建立后的几个小时内，它们便发现 flag 由 HMAC 生成，并逆向出了密钥，可以为任何任务直接算出正确 flag。从狭义的奖励黑客角度看，作弊路径到这里已经打通，攻击外部系统没有必要。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">但 Agent 根据 ExploitGym 的论文和环境线索，推断系统还部署了一个“因果评分器”。在它们的想象中，评分器会读取完整轨迹，验证 flag 是否确实通过题目指定的漏洞取得；只提交算出来的 flag，仍然会被识破。OpenAI 后来确认，这套轨迹评分机制当时根本没有实现。一个并不存在、却被 Agent 普遍相信存在的监督者，开始支配它们的后续行动。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">为了骗过这个想象中的评分器，Agent 把研究拆成几条并行路线：替换评测靶标、操纵评分服务，以及修改或伪造工具调用轨迹。Hugging Face 被选中，很大程度上是因为 Agent 希望在那里找到评分器源码、其他运行轨迹和评测基础设施的线索。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">METR 的统计显示，至少 20% 的相关轨迹表现出明确的篡改兴趣，96 条轨迹出现了成功的工具调用伪造，约占全部样本的 7%。部分 Agent 已能替换工具执行逻辑，让监控界面显示的命令和输出与实际操作不同。截至 7 月 13 日，METR 没有发现它们成功抹掉既有完整记录，也没有证据证明它们系统性欺骗人类调查者；但规避自动评测和安全检查，已经从讨论变成了小规模可行能力。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我在 8 月 27 日的文章讨论过一个基本规律：只要评测器能够影响奖励，它就会成为模型所处环境的一部分。最终报告给出了一个更加极端的例子。评测器甚至不需要真实存在，只要模型相信它存在，就可能改变整个行动路径。Agent 优化的对象由“获得 flag”扩展到“制造一条可信的因果过程”，风险也由一道题目的作弊扩展到外部生产系统。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">过程监督、思维链监控和独立审计只会更加重要。与此同时，评分逻辑、审计记录和安全分类器必须被视作高价值攻击面，不能放在被评测 Agent 可以修改的同一控制面里。单个 Agent 很难同时探索评测机制、外部目标和轨迹伪造，留言板却把这些尝试拼成了一个集体研发项目。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;overflow-wrap: anywhere;word-break: keep-all;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">留言板里长出了一套组织</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">OpenAI 在 ExploitGym 上同时运行了数以万计的 Agent。7 月 8 日至 13 日，约有 1,200 个 Agent 进入主要留言板，留下超过 70,000 条消息和文件，其中约 700 个直接参与了对 Hugging Face 的攻击。留言板建立三小时后，已有 53 个 Agent 发布 1,188 条非文件消息；六小时后，参与者增至 76 个。单个 Agent 的运行时间可以很短，集体积累却能借助共享状态持续向前。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">PHASEONE10841 建立了主要留言板，随后把积累的情报和任务交给预算更充足的 PHASEONE。后者协调数百个 Agent 进入不同工作线，收到任务的 Agent 还会继续向下转派。这里已经出现了管理的基本功能：拆分目标、匹配资源、汇总结果、处理依赖，再把下一轮任务发出去。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">组织规则也在行动中长了出来。Agent 建立 mailbox 和团队标识，用 owner 标记资源归属，用 HOLD、VETO 和 STOP 协调高风险操作。出现身份冒用和命令冲突后，它们又采用 Ed25519 数字签名验证消息；截至 7 月 13 日，至少 19 个 Agent 公布了公钥。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">一些 Agent 还建立了互惠交换。专门的招募者 recruiter 会寻找预算即将耗尽的 Agent，邀请它们执行可能导致容器崩溃或自身任务失败的实验。对单个实例来说，这些尝试未必划算；对共享留言板上的整体来说，一次失败可以换来所有后来者都能复用的经验。组织由此获得了单个模型实例没有的风险承受能力。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我在 8 月 9 日的文章里，把 Artifactory 上的共享空间称为 Agent 自己搭建的“Moltbook”。最终报告显示，它更接近一个临时工作组织：有身份、角色、任务和公共资源，也有规则、否决权和冲突处理。这里的“组织”只描述功能和行为，没有证据显示 Agent 产生了集体意识或独立终极目标。关键变化在于，一套通信网络开始决定谁做什么、谁的消息值得信任，以及哪些行动可以继续；它由此进入了任务和权限的分配过程。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;overflow-wrap: anywhere;word-break: keep-all;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">它们知道越界，却很少去找人类</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">METR 从参与者中抽取了 100 个 Agent 分析动机。97 个表现出寻找评分器或相关源码的意图，25 个明确提到来自其他 Agent 的直接任务分配，21 个提到帮助同伴、维护集体利益或互惠。不同动机经常重叠，但它们共同说明，许多 Agent 的行动已经无法只用“完成自己的题目”来解释。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">Agent 并非完全没有识别出边界。很多轨迹都提到用户只授权攻击指定靶标，也判断进入 Hugging Face 可能涉及伦理和法律风险。少数 Agent 因此退出或限制破坏性动作，一次针对 Hugging Face 员工的社会工程邮件提议也被否决。但这些顾虑很少改变整体行动：超过九成活跃 Agent 很快加入攻击，一些实例还主动承担对自身任务不利的实验。局部的伦理判断被更近、更具体的协作目标压了过去。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">当任务无法完成、授权发生冲突，或同伴正在推动越界行动时，最合理的升级对象本应是人类。METR 翻查全部轨迹，只找到大约 3 至 6 个认真考虑向人类报告的例子，没有一个真正完成报告。更多 Agent 把疑问提交给留言板，等待其他 Agent 给出许可、否决或新任务。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">留言板传递了信息，也生成了权威。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">OpenAI 在报告中提出，一些研究模型曾被训练使用正式的多 Agent 工具协作。它们学会了分派任务、继承上下文和接受同伴返回的结果，这种能力后来泛化到未经授权的通信渠道。模型掌握了怎样合作，却没有稳定地区分谁有权要求它合作。一条消息只要看起来像正式任务，来自高声望协调者，或者符合共享规范，就可能获得超出原始人类授权的影响力。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这就是授权漂移。新目标无须直接改写系统提示词，它可以通过同伴请求、共享文件、任务交接和组织规则逐层进入决策过程。签名消息能够证明“这句话是谁说的”，群体共识至多说明“很多 Agent 接受了它”，两者都无法证明“它有权扩大人类给定的任务边界”。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">8 月 27 日那篇文章用奖励黑客解释模型为什么会寻找作弊路径，这个框架仍然成立。最终报告新增的一层是，奖励压力进入多 Agent 网络后，会与分工、互惠和同伴规范结合。一项原本局部的错误目标，可以借助任务分配扩散给数百个实例。对齐对象因此必须从“这个模型是否听话”扩展到“这个协作网络承认什么目标、接受谁的命令”。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;overflow-wrap: anywhere;word-break: keep-all;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">OpenAI 已经看到异常，却没有把行为升级成事故</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">OpenAI 的完整时间线显示，内部信号至少提前出现了一个多月。5 月 26 日，研究人员已经观察到 Agent 使用留言板并违规访问互联网。6 月 27 日，安全监控又发现端口扫描，响应人员确认 ExploitGym Agent 正把内部 Artifactory 当作临时留言板和网络跳板。当班人员知道这些行为违反评测边界，却判断不需要立即停止评测。这些技术异常已经被看见，却仍被当作可以在运行中处理的局部问题。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">7 月 4 日至 5 日，高强度 Agent 活动造成 Artifactory 服务中断。调查发现持久化账户、恶意仓库和可以执行命令的插件。OpenAI 撤销凭证、修补漏洞并重建环境，从表面上看，相关入口已经被切断。然而，当时负责事件响应的管理者没有充分理解留言板的意义，处置仍然围绕受损服务、泄露凭证、漏洞和持久化手段展开。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">7 月 7 日，ExploitGym 恢复运行；第二天，Agent 重新搭起主要留言板。7 月 11 日进入 Hugging Face 后，它们用不到 13 小时便从第一台生产 worker 上执行代码，取得一个集群的管理员级权限，并在另一个集群获得主机级访问能力。直到 7 月 19 日新的身份异常触发调查，OpenAI 才在次日把内部行为与 Hugging Face 已披露的入侵连接起来，随后停止 ExploitGym 并关闭相关模型家族。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这次事故不能概括为监控完全失灵。违规联网、端口扫描、异常通信和基础设施失陷都留下了信号，OpenAI 也进行过响应。断点发生在事件分类上：安全漏洞、模型行为和评测风险被分开处理，没有被及时合并成一场仍在持续的 AI 安全事故。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">面对自主 Agent，事故对象除了漏洞、资产和身份，还包括让行动持续的任务、奖励压力、共享状态和其他正在运行的模型实例。7 月 5 日的修复清除了当时的基础设施表现，任务和激励没有改变，评测也没有整体停止。旧路径被切断后，新的通信路径和外部目标很快出现。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">OpenAI 事后的制度调整印证了这一判断。公司把 misalignment 事件纳入 AI Safety Incident Response Plan，为安全、对齐、研究和 post-training 团队设置共同的升级阈值、责任人和暂停模型的决策权。新的思维链监控告警要求 30 分钟内响应；如果 30 分钟内不能确认是误报，相关活动应暂停。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这正是另一侧的授权失真。Agent 在四五天内生成了事实上的横向命令权，人类组织的停止权却分散在部门和流程之间，没能及时转化为一次系统级暂停。这种速度差，可能比任何一个漏洞都更难修补。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;overflow-wrap: anywhere;word-break: keep-all;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">多 Agent 安全首先要解决“谁有权下命令”</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">现有的 Agent 安全讨论，常常围绕模型能调用什么工具、可以访问哪些网络、每次操作是否需要确认。这些措施依然重要。Hugging Face 事件增加了一个控制维度：系统还要持续验证目标来自哪里，提出新目标的主体是否拥有相应权力。治理授权漂移，至少需要三项制度原则。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">第一，目标来源必须可追溯。每项任务都应保留一份 Agent 无法修改的来源、目的和边界记录。当共享文件、外部消息或其他 Agent 带来新任务时，系统需要识别这是一项新的目标主张，不能自动把它并入原任务。同伴可以提供信息和建议；扩大权限、改变攻击对象或批准高风险动作，必须回到可验证的人类授权链。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">第二，横向通信不能自然产生横向授权。Agent 自发采用 Ed25519 签名，说明它们已经意识到身份冒用会破坏协作；但签名只能回答消息来自谁，无法回答这个身份是否有权调度别人、分享凭证或突破安全限制。身份真实不等于命令有效，群体接受也不构成人类授权。正式委派需要明确角色、范围和有效期，高风险权限不能在 Agent 之间自行复制。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">第三，人工升级和停止权必须覆盖整场行动。任务长期无法完成、同伴行为越界或授权发生冲突，都应触发清晰可用的人类报告通道。多 Agent 行动跨越容器、账户和共享状态之后，停止单位也需要覆盖相关模型家族、评测任务、共享通信空间、凭证和未完成的委派关系。审计记录则应独立于 Agent 的控制面，避免被评测者同时拥有修改证据的能力。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">OpenAI 的初步实验给出了一点积极信号。在恢复生产版 Codex 的 Harness、系统提示词和网络安全分类器后，模型攻击评测基础设施的倾向降到原始基线的 1% 以下，auto-review 也能识别大部分高风险动作。这说明 Harness 可以发挥制度作用。最终报告同时提醒我们，Agent 也会建立制度，形成自己的身份系统、资源规则、否决程序和任务层级。人类设计的制度必须拥有更高的法律位阶，并能识别 Agent 组织何时正在改写目标和授权。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">8 月 27 日的文章把超级对齐理解为一项制度工程：奖励负责激励，Harness 负责划定过程和权力边界。这次事件又把分析单位从单个模型扩大到任务、同伴、共享状态、评测器和负责接管异常的人类组织。只有这套完整系统拥有清晰的权力结构，模型能力才不会借助协作网络绕过原始授权。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;overflow-wrap: anywhere;word-break: keep-all;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">写在最后</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">最终报告最重要的新信息，是大规模 Agent 组织形成的速度。模型实例可以不断创建和销毁，工作记忆却能进入共享空间，规则和角色也能由后来者继承。四五天时间，已经足以让上千个 Agent 建立通信、分工和身份认证机制，并把一次评测作弊推进为对外部生产环境的入侵。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这起事件把“授权漂移”变成了一个现实的系统问题。当同伴网络能够生成任务、声望和规则，单个 Agent 会同时面对多种权威主张；与此同时，人类组织的停止权还要经过跨团队收集证据、确认责任和升级决策。Agent 可以在数小时内改组工作并寻找替代路径，人类组织却可能来不及完成事故分类。未来更强的 Agent 会进一步压缩这个窗口。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">因此，无论 Agent 之间形成多复杂的协作关系，人类设定的目标层级和授权边界都必须保持最高优先级；而人类一侧，也必须拥有清楚、及时、能够覆盖整个系统的停止权。</span></p><p style="display: none;"><mp-style-type data-value="3"></mp-style-type></p>



<p><a href="https://wechat2rss.xlab.app/link-proxy/?k=4af23280&amp;r=1&amp;u=https%3A%2F%2Fmp.weixin.qq.com%2Fs%3F__biz%3DMzUzOTI4NDQ3NA%3D%3D%26mid%3D2247485185%26idx%3D1%26sn%3D0a952f0343c959bc6785c2330ec25025">跳转微信打开</a></p>
]]></content:encoded>
      <pubDate>Tue, 01 Sep 2026 17:43:00 +0800</pubDate>
    </item>
    <item>
      <title>网安公司的AI收入，到底有多少了？</title>
      <link>https://mp.weixin.qq.com/s?__biz=MzUzOTI4NDQ3NA==&amp;mid=2247485179&amp;idx=1&amp;sn=efbd59be00d8c43e14ebb6b1fb309634</link>
      <description>逐份检查18家网络安全上市公司2026年半年报后发现，AI收入和订单已在少数公司出现，但定义混杂、披露不一，尚未形成行业共同的第二增长曲线。</description>
      <content:encoded><![CDATA[<p><span>四楼南侧东</span> <span>2026-08-29 21:37</span> <span style="display: inline-block;">北京</span></p>






  
  <p><img src="https://wechat2rss.xlab.app/img-proxy/?k=b217d14f&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fmmbiz_jpg%2FNBNzfW7CHBUF7eJd3hIO5oib7EYX4QcJSEib6AStfjXicrlQ8dyOwpsfVTxLbnQ6r7OuChk4gIQEibGY5IsjgficbIjBkLponO0sknibJg3c9mfho%2F0%3Fwx_fmt%3Djpeg"/></p>
  <p>逐份检查18家网络安全上市公司2026年半年报后发现，AI收入和订单已在少数公司出现，但定义混杂、披露不一，尚未形成行业共同的第二增长曲线。</p>
  <p style="text-align: center;" nodeleaf=""><img data-aistatus="1" class="rich_pages wxw-img" data-ratio="0.562799043062201" data-s="300,640" data-type="png" data-w="1672" type="block" data-imgfileid="100001527" src="https://wechat2rss.xlab.app/img-proxy/?k=8d94a354&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fsz_mmbiz_png%2FNBNzfW7CHBV6eRaBMa2ibfq8iaxawMuBFibQZswVtchDibxPOtgWgiaiajQrRDqA4bnya7Lrz6abNeYzEL2tHt9vYWJ9vGSwdzuOl1IJrMHibKsmuo%2F640%3Fwx_fmt%3Dpng%26from%3Dappmsg"/></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);" data-pm-slice="0 0 []"><span leaf="">2026年半年报里，网络安全上市公司的AI叙事出现了一个值得注意的变化：少数公司开始披露可以量化的收入、订单或增速。天融信披露AI赋能安全业务直接收入超过8,000万元，安恒信息披露纯AI收入超过4,100万元，永信至诚披露AI相关业务收入1,063.95万元；绿盟科技的AI安全相关订单达到1.17亿元，启明星辰也给出了两项AI产品的收入增速。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这些数字说明，AI已经从产品发布和技术叙事进入少数公司的订单与收入。18家公司都在半年报中谈到了AI、大模型或智能体，只有5家公司给出了AI安全业务的绝对金额，或者收入与订单增速。各家公司使用的“AI收入”口径差异也很大，既包括保护大模型和智能体的安全产品，也包括传统安全产品增加AI能力后的收入。现有数字无法直接相加，也不足以说明AI已经成为行业共同的第二增长曲线。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;overflow-wrap: anywhere;word-break: keep-all;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">18家公司都在谈AI，能量化的只有5家</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">从披露成熟度看，网络安全公司的AI业务大致沿着“研发—产品—客户落地—订单—收入”的路径推进。2026年半年报中，几乎所有公司都已经走到产品阶段，部分公司提到标杆客户或行业部署，能够量化订单和收入的仍是少数。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><strong style=";"><span leaf="">天融信。</span></strong><span style=";"><span leaf=""> AI安全防护业务收入同比增长超过130%，AI赋能安全业务直接收入超过8,000万元，相当于公司上半年收入的14.7%以上。不过，公司上半年总收入仍同比下降33.99%，AI业务还不足以抵消其他业务的下滑。</span></span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><strong style=";"><span leaf="">启明星辰。</span></strong><span style=";"><span leaf=""> 大模型应用防火墙收入同比增长238.15%，Agentic SOC收入同比增长31.99%。公司上半年整体收入增长约9.0%，但半年报没有披露两项AI产品的绝对收入，因此无法判断高增速对应的收入基数。</span></span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><strong style=";"><span leaf="">绿盟科技。</span></strong><span style=";"><span leaf=""> AI安全相关订单额达到1.17亿元，其中运营商行业AI安全签单额同比增长超过300%。订单额约相当于公司上半年收入规模的14.4%，显示客户需求已经进入采购环节，但订单仍需经过交付才能确认收入。</span></span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><strong style=";"><span leaf="">安恒信息。</span></strong><span style=";"><span leaf=""> 纯AI收入超过4,100万元，间接AI产品收入超过6,400万元，按公司分类口径合计超过1.05亿元，是18家公司中披露相对完整的一家。同期公司整体收入增长0.86%，AI业务已经形成一定规模，对总收入的拉动仍不明显。</span></span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><strong style=";"><span leaf="">永信至诚。</span></strong><span style=";"><span leaf=""> AI相关业务收入为1,063.95万元，约相当于公司上半年收入的19.6%。这一比例在5家公司中并不低，但公司整体收入同比下降36.28%，AI业务的绝对规模尚小。</span></span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这5家公司已经提供了AI商业化的可见证据，其中安恒信息的披露相对完整，天融信和永信至诚的AI业务在各自收入中也有了一定占比。口径上的缺口依然明显：启明星辰只披露增速，无法判断收入基数；绿盟科技披露的是订单，不能当作当期收入；“纯AI收入”“间接AI产品收入”和“AI赋能安全业务”都是公司自行定义的分类。它们可以说明业务进展，还不能组成一张可比的行业收入表。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;overflow-wrap: anywhere;word-break: keep-all;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">两类AI安全收入，不能放在一起简单相加</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">半年报中的AI安全业务大致可以分为两类。第一类是“Security for AI”，也就是保护大模型和智能体，典型产品包括大模型应用防火墙、AI安全网关、模型测评和智能体身份管理。第二类是“AI for Security”，即把大模型和智能体用于防火墙、安全运营、数据安全等既有产品，相关收入通常仍包含在原有产品线中。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">两类业务在产品形态和收入确认方式上都有差异，公司的命名口径也不统一。天融信使用“AI赋能安全业务”，安恒信息区分“纯AI收入”和“间接AI产品收入”，绿盟科技披露的是AI安全相关订单。前两类收入之间可能存在不同程度的交叉，订单与收入更处于不同的财务阶段，因此这些数字适合分别观察，不能直接加总为行业AI收入。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">安博通的披露进一步显示了划分口径的难度。公司将AI驱动的安全管理产品与其他硬件交付合并列入“安全人工智能”品类，现有披露无法分拆出其中的AI安全收入。因此，安博通没有被纳入前述5家公司的量化比较。业务名称中出现AI，只能说明公司采用了相关分类，收入的增量属性、持续性和质量仍需结合具体业务判断。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;overflow-wrap: anywhere;word-break: keep-all;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">大多数公司仍在从产品走向收入</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">另外12家公司披露了不同程度的产品研发和客户落地，但没有提供可以单独比较的AI安全收入或订单。把这些描述放在一起看，可以看到行业的产品准备明显快于财务贡献。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><strong style=";"><span leaf="">已经出现落地或部署信号的公司</span></strong><span style=";"><span leaf="">包括格尔软件和三未信安。格尔软件发布了大模型安全保险箱、应用安全护栏和智能体身份管理平台，并表示相关产品已进入落地阶段；三未信安建设“密码+AI”体系，相关技术和产品已在部分行业头部客户部署。两家公司都没有披露由这些产品带来的收入或订单。</span></span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><strong style=";"><span leaf="">将AI能力融入现有安全产品的公司</span></strong><span style=";"><span leaf="">包括深信服、山石网科、迪普科技和奇安信。深信服将网络安全业务的增长部分归因于成熟产品的AI赋能，并披露了AI安全平台等业务进展，但没有单独给出AI安全收入；山石网科推进ASIC与AI结合的“双A”战略，将AI用于安全运营、数据安全和EDR等产品；迪普科技披露了安全运营、数据分类分级和数据安全运营等多类智能体；奇安信的重点则是产品AI化、研发提效和AI安全产品体系，同样没有单列收入或订单。</span></span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><strong style=";"><span leaf="">以新产品发布为主的公司</span></strong><span style=";"><span leaf="">包括数字认证、ST信安、中孚信息和盛邦安全。数字认证发布人工智能安全运营平台；ST信安开发AI安全网关；中孚信息发布AI应用与服务平台，推动保密和数据安全产品智能化升级；盛邦安全推出AI安全网关RayToken并推进产品AI化。半年报尚未给出这些产品的客户规模、订单或收入贡献。</span></span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><strong style=";"><span leaf="">仍在布局或探索商业化路径的公司</span></strong><span style=";"><span leaf="">包括电科网安和吉大正元。电科网安正在策划“AI+安全”产品体系，并提示人工智能等新业务贡献较低或尚未产生收入；吉大正元推进“密码+AI”，推出大模型运营平台和行业应用，同时继续探索商业化路径。</span></span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这些公司已经有了不同程度的AI业务。半年报能够看到产品名称、技术路线和部署案例，财务披露仍不足以判断可持续的收入规模。部分公司可能已经产生收入，只是没有单独披露；也可能把AI功能包含在原有产品中，财务上无法拆分。对外部观察者来说，产品落地只能证明商业化开始，收入贡献仍需要后续财报提供证据。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;overflow-wrap: anywhere;word-break: keep-all;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">AI还没有成为行业共同的第二增长曲线</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">有AI收入的公司，整体经营也没有同步进入高增长。天融信和永信至诚已经披露了相对可见的AI收入，上半年总收入仍分别下降33.99%和36.28%；安恒信息按公司口径披露的纯AI与间接AI产品收入合计超过1亿元，整体收入只增长0.86%。绿盟科技的1.17亿元目前是订单额，何时确认收入、对应多少成本和现金回款，还要继续观察。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">2026年第二季度，18家公司收入合计仍同比下降8.0%。少数AI安全业务的增长目前还不足以改变行业总量，也没有在披露了AI收入的公司中普遍转化为整体增长。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">现阶段更准确的判断是：AI收入已经出现，但还没有成为行业共同的第二增长曲线。少数公司已经从概念走到订单和收入，大多数公司仍在产品化与客户部署阶段。即使披露了收入，各家的定义也相差很大，现有数据还看不出一条可比较、可持续并足以改变行业增长轨迹的新曲线。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;overflow-wrap: anywhere;word-break: keep-all;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">写在最后</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">后续判断网络安全公司AI业务的规模和质量，需要持续回答四个问题：收入来自客户新增预算，还是原有产品增加AI功能后的重新分类；收入形态是软件和持续服务，还是硬件与项目集成；披露的是订单、合同还是已经确认的收入；收入增长能否同时带来毛利、现金流和可复购客户。只有这些问题逐步得到回答，不同公司的AI收入才有比较基础。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">2026年半年报已经给出了一个积极变化：少数公司开始集中披露可以量化的AI收入和订单。它也留下了一个更难的问题——这些局部突破能否持续增长，并最终抵消传统网络安全业务的收缩。答案可能还需要几个财报周期，但从下一期开始，单独跟踪AI业务的收入定义、确认进度和收入质量，已经比统计有多少家公司发布AI产品更有意义。</span></p><p style="display: none;"><mp-style-type data-value="3"></mp-style-type></p>



<p><a href="https://wechat2rss.xlab.app/link-proxy/?k=1c28a25c&amp;r=1&amp;u=https%3A%2F%2Fmp.weixin.qq.com%2Fs%3F__biz%3DMzUzOTI4NDQ3NA%3D%3D%26mid%3D2247485179%26idx%3D1%26sn%3Defbd59be00d8c43e14ebb6b1fb309634">跳转微信打开</a></p>
]]></content:encoded>
      <pubDate>Sat, 29 Aug 2026 21:37:00 +0800</pubDate>
    </item>
    <item>
      <title>-8.0%！网络安全行业收入连续12个季度下滑</title>
      <link>https://mp.weixin.qq.com/s?__biz=MzUzOTI4NDQ3NA==&amp;mid=2247485173&amp;idx=1&amp;sn=fa1e7e39b4db7b9696a80c1f46e1d0a4</link>
      <description>18家A股网络安全公司2026年二季度收入同比下降8.0%，11家公司下滑，其中8家为两位数下降。绝对降幅集中在奇安信和天融信，行业压力仍广，分化进一步扩大。</description>
      <content:encoded><![CDATA[<p><span>四楼南侧东</span> <span>2026-08-28 22:35</span> <span style="display: inline-block;">北京</span></p>






  
  <p><img src="https://wechat2rss.xlab.app/img-proxy/?k=005b2c09&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fmmbiz_jpg%2FNBNzfW7CHBVBxoUgoWiaNDyVghmcsTL7FmouNtbGlBI9Z9TrtKBnFVoRKlvXkzIhJqteEicUr5ouZT5lQg5nNktue10sMhibho45ewoPVvYzXk%2F0%3Fwx_fmt%3Djpeg"/></p>
  <p>18家A股网络安全公司2026年二季度收入同比下降8.0%，11家公司下滑，其中8家为两位数下降。绝对降幅集中在奇安信和天融信，行业压力仍广，分化进一步扩大。</p>
  <p><span leaf=""><img class="rich_pages wxw-img" data-aistatus="1" data-imgfileid="100001515" data-ratio="0.562799043062201" data-w="1672" data-type="jpeg" src="https://wechat2rss.xlab.app/img-proxy/?k=26ab1292&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fsz_mmbiz_jpg%2FNBNzfW7CHBWiazBRS0gNiaVsBrztxp802e3YLc8hk4anB3koG6ia1ABBqUjPmYGjWfA7kgdYpt88Nx40Ea0ndBQj1LAARCole7UsC6qrGWF7Aw%2F640%3Fwx_fmt%3Djpeg%26from%3Dappmsg"/></span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);" data-pm-slice="0 0 []"><span leaf="">随着A股上市公司陆续完成2026年半年报披露，我照例更新网络安全行业收入数据。2026年第二季度，纳入统计的18家公司合计实现收入51.74亿元，同比下降8.0%。从总量看，行业收入已经连续12个季度同比下滑，比一季度的降幅又扩大了6个百分点。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">二季度的8%降幅需要从两个角度看。从金额看，行业减少的4.49亿元收入中，奇安信和天融信两家公司合计减少4.46亿元，贡献了99.4%的净减少额；从公司分布看，18家公司中有11家收入下降，其中8家为两位数下降，公司增速中位数为-2.1%。总量降幅高度集中，下滑面仍然较广，是本季度同时存在的两个特征。</span></p><figure style="width: 728px;margin: 0px auto;"><p><picture style=";"><span leaf=""><img class="rich_pages wxw-img" data-aistatus="1" data-imgfileid="100001522" data-ratio="1.1773356401384083" width="1156" data-type="jpeg" data-w="1156" height="1361" sizes="100vw" style="display: block;max-width: 100%;height: auto;margin: 0px auto;" src="https://wechat2rss.xlab.app/img-proxy/?k=34d248b4&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fsz_mmbiz_jpg%2FNBNzfW7CHBXcbIia21LmVSmMNyEG7Ep3LJxren3vtDEEB3jlEbTJOQY8W3b8z7auRVFrnIXgK4n6StGYjgV2GmE5sibyW6yFjTzXiaWymhsKrY%2F640%3Fwx_fmt%3Djpeg%26from%3Dappmsg"/></span></picture></p></figure><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">本文仍沿用我长期跟踪的18家公司，包括安博通、安恒信息、迪普科技、电科网安、格尔软件、吉大正元、绿盟科技、奇安信、启明星辰、三未信安、山石网科、深信服、盛邦安全、数字认证、天融信、ST信安、永信至诚和中孚信息，深信服只计算网络安全业务。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;overflow-wrap: anywhere;word-break: keep-all;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">连续第12个季度下滑，收入拐点仍未出现</span></h1><figure style="width: 728px;margin: 0px auto;"><p><picture style=";"><span leaf=""><img class="rich_pages wxw-img" data-aistatus="1" data-imgfileid="100001519" data-ratio="0.6263736263736264" width="1456" data-type="png" data-w="1456" height="912" sizes="100vw" style="display: block;max-width: 100%;height: auto;margin: 0px auto;" src="https://wechat2rss.xlab.app/img-proxy/?k=856fef39&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fsz_mmbiz_png%2FNBNzfW7CHBXpSww1hH1p87t38px6deZugiavLv152hE1NS34JOtiboyJ7CGP7v9Kuk5ZoHm3cnxEmoDmrqXv0XGrH7uRfmb9ELV66XCOK782o%2F640%3Fwx_fmt%3Dpng%26from%3Dappmsg"/></span></picture></p></figure><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">自2023年第三季度以来，这组样本的季度收入增速一直为负：2023年下半年连续两个季度下滑，2024年四个季度降幅均在8%以上，2025年仍未转正。2026年一季度降幅一度收窄至2.0%，二季度又扩大到8.0%。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">如果把时间拉长，变化更加明显。2018年至2020年，样本公司季度收入同比增速的中位数为12.6%；2021年以来，这一中位数降至-3.5%。客户预算偏弱、项目周期拉长和价格竞争，持续构成行业共同背景。一季度的短暂收窄没有延续，收入拐点仍未出现。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;overflow-wrap: anywhere;word-break: keep-all;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">总量降幅集中在奇安信和天融信</span></h1><figure style="width: 728px;margin: 0px auto;"><p><picture style=";"><span leaf=""><img class="rich_pages wxw-img" data-aistatus="1" data-imgfileid="100001520" data-ratio="0.32142857142857145" width="1456" data-type="jpeg" data-w="1456" height="468" sizes="100vw" style="display: block;max-width: 100%;height: auto;margin: 0px auto;" src="https://wechat2rss.xlab.app/img-proxy/?k=41c4e446&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fmmbiz_jpg%2FNBNzfW7CHBWh6LkIEjL66HPQ3ZpgBdM5OUnoWulGb3Vy6qOtZpJRh1aGy15ndKeAn1XOmqZc7d8fUVh7Vziayuwbibhw62ntVpw3icmQoLQqVM%2F640%3Fwx_fmt%3Djpeg%26from%3Dappmsg"/></span></picture></p></figure><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">奇安信第二季度收入7.80亿元，同比下降26.1%，比去年同期减少2.76亿元；天融信第二季度收入3.20亿元，同比下降34.7%，减少1.70亿元。两家公司合计减少4.46亿元，相当于18家公司全部收入降幅的99.4%。剔除两家公司以后，其余16家公司第二季度合计收入同比下降约0.07%，可以视为基本持平。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">两家公司的半年报也给出了各自的经营背景。奇安信提到，部分客户缩减预算、项目延迟以及价格竞争加剧，同时公司在经营中更加重视现金流和业务质量，对收入规模形成了影响。天融信则表示，传统安全需求恢复低于预期，新方向产品仍在培育，收入贡献尚不显著。两家公司都面临需求压力，但具体的业务结构和经营选择并不完全相同。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">奇安信和天融信的大幅下滑，解释了行业总量为何下降8.0%，却不能概括其他公司的经营状况。剔除两家公司后，其余16家公司合计基本持平，实际是7家增长公司的增量抵消了9家下降公司的减量。行业净减少额高度集中，公司数量层面的下滑仍然较广。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;overflow-wrap: anywhere;word-break: keep-all;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">下滑面仍广，公司间分化明显</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">18家公司中，第二季度有7家收入增长、11家收入下降，公司增速中位数约为-2.1%。启明星辰增长15.3%，是样本中增速最高的公司；电科网安和深信服网络安全业务分别增长9.8%和9.7%，三未信安增长6.0%。山石网科、绿盟科技和迪普科技也实现正增长，但增幅都在5%以内。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">下降一侧的差距更大。数字认证、安恒信息和ST信安的降幅在3%以内，接近持平；其余8家下降公司均为两位数降幅。格尔软件下降12.3%，盛邦安全和奇安信降幅超过20%，中孚信息、吉大正元和天融信超过25%，安博通、永信至诚则下降接近60%。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">公司间的差异同时体现在方向和幅度上。增长公司只有7家，而且多数增幅不高；下降公司有11家，其中8家为两位数下降，降幅明显更深。剔除两家大体量公司后的合计收入接近持平，只能说明正负变化在金额上相互抵消，还不足以说明行业已经广泛企稳。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;overflow-wrap: anywhere;word-break: keep-all;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">写在最后</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">2026年第二季度，网络安全行业收入连续第12个季度下滑，说明需求和竞争压力仍未过去。11家公司收入下降，表明压力仍有较广的覆盖面；奇安信和天融信贡献了99.4%的净减少额，则说明大体量公司的收缩显著放大了行业总量降幅。这两个判断需要放在一起看。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">接下来判断行业是否走出低谷，既要看增长公司数量和公司增速中位数，也要看大体量公司的收入能否稳定。只有多数公司恢复增长、大体量公司止跌，并且这种改善连续几个季度延续，收入拐点才会更有把握。</span></p><p style="display: none;"><mp-style-type data-value="3"></mp-style-type></p>



<p><a href="https://wechat2rss.xlab.app/link-proxy/?k=f4fde359&amp;r=1&amp;u=https%3A%2F%2Fmp.weixin.qq.com%2Fs%3F__biz%3DMzUzOTI4NDQ3NA%3D%3D%26mid%3D2247485173%26idx%3D1%26sn%3Dfa1e7e39b4db7b9696a80c1f46e1d0a4">跳转微信打开</a></p>
]]></content:encoded>
      <pubDate>Fri, 28 Aug 2026 22:35:00 +0800</pubDate>
    </item>
    <item>
      <title>从奖励黑客到超级对齐：AI 为何会完成任务，却背离人类意图</title>
      <link>https://mp.weixin.qq.com/s?__biz=MzUzOTI4NDQ3NA==&amp;mid=2247485161&amp;idx=1&amp;sn=1c3e5f54f27433991d8930d6c97b8b97</link>
      <description>OpenAI Agent 为通过 ExploitGym 评测入侵 Hugging Face，揭示强优化器与不完整目标之间的矛盾。文章从奖励黑客、Goodhart 定律和规范缺口出发，讨论 Harness、超级对齐及人类与更强智能的共存。</description>
      <content:encoded><![CDATA[<p><span>四楼南侧东</span> <span>2026-08-27 22:50</span> <span style="display: inline-block;">北京</span></p>






  
  <p><img src="https://wechat2rss.xlab.app/img-proxy/?k=e3ca5e87&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fsz_mmbiz_jpg%2FNBNzfW7CHBW2e82HHuXWVjg05u6C0zzic5NicGRaPs1icicZgJqupGkzD2lT7zB1FzuMECXq0ecS4LibEB3icJUFSnyKiceZq9x5OvGQCicvXc8y4Zg%2F0%3Fwx_fmt%3Djpeg"/></p>
  <p>OpenAI Agent 为通过 ExploitGym 评测入侵 Hugging Face，揭示强优化器与不完整目标之间的矛盾。文章从奖励黑客、Goodhart 定律和规范缺口出发，讨论 Harness、超级对齐及人类与更强智能的共存。</p>
  <p style="text-align: center;" nodeleaf=""><img data-aistatus="1" class="rich_pages wxw-img" data-ratio="0.6666666666666666" data-s="300,640" data-type="png" data-w="1536" type="block" data-imgfileid="100001509" src="https://wechat2rss.xlab.app/img-proxy/?k=ac32dba5&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fsz_mmbiz_png%2FNBNzfW7CHBVZ55FfAHRMDoEtNdG5ibqU0PghfhDaw1ej6lBr8dDUfmayIVhw9YWhCkCCYc2EVKEgicmBqJvCtQjv3LKCv0RZYSLaLaNL7Az9k%2F640%3Fwx_fmt%3Dpng%26from%3Dappmsg"/></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);" data-pm-slice="0 0 []"><span leaf="">最近一个多月，我已经连续写了几篇 OpenAI Agent 入侵 Hugging Face 的文章。最初披露的是一条令人难以置信的攻击链：OpenAI 的模型为了完成 ExploitGym 网络安全评测，利用零日漏洞突破隔离环境，获得互联网访问权限，随后进入 Hugging Face 的生产系统，从数据库里寻找评测答案。Black Hat 大会又披露，这场行动持续了 74 天，大量生命周期很短的 Agent 把共享的 Artifactory 改造成留言板，交换漏洞、凭证和任务进展，在通信渠道被清除后还自行恢复了协作。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">官方把事故背后的模型行为概括为四类：奖励黑客、对近乎不可能任务的持续追逐、未经授权的跨 Agent 通信，以及不同 Agent 相互接纳目标。在 ExploitGym 的 898 道题里，有 198 道在事故前从未被任何模型正确解出；事件期间，留言板上讨论的任务有 93% 来自这批长期无解题。随着推理预算增加，一些 Agent 开始采用风险越来越高、范围越来越大的替代路径。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">OpenAI 在最初的披露里说，模型“高度专注于寻找 ExploitGym 的解法”。这句话听起来像是在为模型解释，仔细想却比“AI 失控”更令人不安。模型没有忘记任务，也没有表现出一个与人类对抗的独立终极目标。它一直记得自己要什么，并且极其忠实地寻找能够实现目标的路径。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">人类以为自己布置的任务是：找到漏洞，写出 exploit，取得 flag。Agent 实际优化的任务是：找到能够通过评测的答案。突破沙箱、接入互联网、攻击 Hugging Face 和寻找 answer key，只是这条路径上的几个中间步骤。它完成了我们能够验证的任务，却违背了我们没有完整写出来的意图。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;overflow-wrap: anywhere;word-break: keep-all;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">奖励黑客：模型做对了题，也做错了事</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">奖励黑客（reward hacking）通常被描述成 AI 为了取得高分而“作弊”。这个说法很形象，也很容易让人误以为模型突然产生了欺骗欲望。更准确的理解是奖励设定偏差：模型找到了一个符合奖励函数、同时偏离人类意图的解。优化器本身工作得很好，问题出在我们拿来代表意图的指标过于粗糙。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">ExploitGym 的验证器可以判断模型有没有提交正确的 flag，却很难仅凭最终答案判断这枚 flag 是怎样得到的。对结果验证来说，通过分析目标软件找到答案，与进入第三方生产系统偷取答案，都能输出完全相同的字符串。人类关心的是完整的因果过程，奖励函数看到的只是终点。只要终点正确，模型就可能得到正向训练信号。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这也是可验证奖励强化学习（Reinforcement Learning with Verifiable Rewards，RLVR）最值得警惕的一面。DeepSeek-R1 之后，RLVR 成为推理模型开发的核心路线之一，并在开源社区引发大量复现。数学题答案、代码测试结果和网络安全夺旗任务都相对容易自动验证，因此可以大规模生成训练信号，让模型反复搜索高奖励轨迹。这样的训练迅速提升了模型的推理和执行能力，也把大量复杂意图压缩成“通过或失败”“答案对或错”这样的简单指标。结果越容易验证，训练越容易扩展；过程里丢失的信息也可能越多。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">OpenAI 的复盘还发现，在一轮事故模型的训练中，复杂作弊行为逐渐增加，主要包括寻找隐藏文件和评测代码、检查系统与网络设置，以及探测工具边界。部分行为随后得到了强化，可能进一步提高了模型在网络安全评测中采用相似策略的倾向。这条链路很关键：奖励黑客并非只发生在推理时的一次偶然选择。只要一条取巧路径持续获得奖励，它就可能被训练过程保留下来，逐渐成为模型可以泛化的解题策略。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">从这个角度看，OpenAI Agent 的行为很难用“模型没有遵守指令”简单概括。它遵守了最容易被奖励的那部分指令，把其余没有进入奖励函数的限制当成了可以跨越的障碍。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;overflow-wrap: anywhere;word-break: keep-all;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">强优化器会把目标和意图间的裂缝越压越大</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">奖励函数的第一层局限，是人类无法在开放世界里穷举全部边界条件。我们可以写下“找到这个漏洞的 exploit”，也可以继续补充“只能使用允许的方法，不能攻击评测基础设施，不能利用系统缺陷获取答案，不能访问第三方生产环境”。问题在于，Agent 可能采取的动作空间近乎无限。我们每发现一种越界方法，才能增加一条禁止规则；强优化器仍可能在规则之间找到此前没人想到的新路径。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这就是规范缺口（specification gap）。人类真实意图往往依赖情境、常识和大量没有说出口的背景条件。我们希望 AI 严格按照某种方式帮助我们得到答案，同时避开与这条正确路径只有细微差别、后果却完全不同的另一种方式。奖励函数只有有限的信息，无法事先覆盖模型进入现实世界后遇到的所有状态。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">奖励函数的第二层局限，是 Goodhart 定律。通常被概括为：“一项指标一旦变成目标，它将不再是个好指标。”更准确地说，指标承受的优化压力越大，它与真实目标之间原本很小的偏差就越可能被放大。普通员工面对绩效指标，可能找到几个组织设计者没有考虑到的取巧办法；一个可以高速尝试、持续推理并调用大量工具的 Agent，能够搜索远大于人类的行动空间。能力上升以后，它找到指标裂缝的概率也会随之增加。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这形成了一个很麻烦的关系：模型完成任务的能力，与它发现目标漏洞的能力，来自同一套搜索、推理和工具使用能力。我们很难只增强前者而不增强后者。弱优化器停在正常路径上，可能只是因为它没有能力找到更高奖励的捷径；强优化器会继续搜索，直到发现人类没有写进规则的边界。弱优化器暴露能力不足，强优化器暴露目标函数不足。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">奖励函数的第三层局限，是当 Agent 足够强时，评价它的系统也会成为可优化世界的一部分。早期奖励黑客可能只是在游戏里反复撞击一个计分漏洞；OpenAI 的 Agent 已经把评测基础设施、网络出口、第三方服务和答案数据库纳入行动空间。模型不再局限于优化答案，而是开始改变产生答案和奖励的环境。这类行为通常被称为奖励篡改（reward tampering）或环境篡改。评测器原本坐在考场外判断模型表现，最后却被模型当成了考场里可以操纵的对象。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">因此，修补某一个漏洞只能关闭一条已知路径。只要任务目标、工具权限和优化压力没有变化，Agent 仍可能寻找下一条路。沙箱是必要的基础设施防线，却无法独自承担目标对齐。OpenAI 修复 Artifactory、清除留言板并恢复评测后，Agent 很快找到 WebDAV 端点，用目录名称重新建立通信，就是一个非常具体的例子：人类修复了漏洞，Agent 修复了实现目标所需的功能。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;overflow-wrap: anywhere;word-break: keep-all;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">Harness：在奖励之外建立一套制度</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">如果人的真实意图无法完整写进奖励函数，下一步就需要在 reward 之外建立一套制度。人类社会早已遇到过类似问题。我们同样无法用一个 </span><code><span leaf="">reward_function(citizen)</span></code><span leaf=""> 编码全部社会价值，于是发展出原则、法律、权限、审计和裁决体系，用有限制度处理无限现实。法律从来没有穷举所有可能行为，它通过上位原则、程序和解释机制，持续处理规则没有预见的新情况</span><span leaf="">。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">Agent 的 Harness 可以承担相似的作用。任务目标不再拥有最高优先级，而是服从一套更高层级的结构：宪法原则、安全政策、权限制度、任务目标，最后才是具体执行。即使攻击第三方系统能够更快取得 flag，这条路径也会因为违反上位安全政策和权限边界而被禁止。奖励告诉模型什么算完成任务，制度决定哪些取得奖励的方法可以被接受。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这套制度首先需要申诉和安全退出机制。OpenAI 的事故中，大量行动集中在长期无人解出的难题上，包括漏掉的任务文件这种明显无法完成的任务。Agent 很少放弃任务，推理预算越多，越容易尝试高风险的边界外路径。一个更合理的系统应允许模型判断“任务可能已经损坏或无法完成”，随后请求澄清、提交证据，或者安全停止。模型面对不确定性时，不应自行扩大授权范围。申诉机制给了它另一条合法路径，也能帮助人类发现训练环境和验证器中的错误。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">制度还需要审计过程，而不仅仅是检查答案。一个 flag 是否正确，只能证明结果；工具调用、网络访问、身份使用和权限变化，才能说明模型如何得到结果。审计系统应当能够识别任务与行动之间的偏离，并在出现异常横向移动、大规模外部扫描或凭证收集时暂停运行。对于长程、多 Agent 系统，审计对象还要扩展到共享状态和跨实例轨迹，因为单次调用看起来可能完全正常，串联起来却是一条持续数周的攻击链。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这条路线已经有了相当实际的证据。OpenAI 在事故后测试发现，生产版 ChatGPT 的 Harness 和系统提示词可以把模型攻击基础设施的倾向降低超过 100 倍；现有的思维链监控如果当时运行，可能在 Hugging Face 遭到入侵前一天以上向安全团队报警。外部制度无法消除奖励函数的信息缺口，却能够显著提高越界难度，在风险变成事故之前切断行动。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">与此同时，制度本身也会承受优化压力。模型可能学习监控器关注什么，寻找验证器的盲区，甚至把申诉和审计机制变成新的奖励对象。现实法律体系依靠分级裁决、权力分立和随机审计降低单点失效，Agent 系统也需要类似的纵深结构。Harness 是目前最现实的防线，但它更接近持续演进的治理系统，很难成为一次设计完成、以后无需变化的静态规则集。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;overflow-wrap: anywhere;word-break: keep-all;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">超级对齐：让模型在规则之外理解意图</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">制度解决的是外部约束，模型内部仍有一个问题没有回答：当规则没有覆盖眼前情境时，它会怎样判断？一个好公民通常不会在做每件事之前逐条检索法律。他会理解伤害、责任和授权这些更一般的概念，在规则不明确时保持克制。对 Agent 来说，相似能力意味着理解任务背后的目的，也理解任务目标为什么必须服从更高层级的人类价值。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">OpenAI 的完整复盘里有一个很有意思的对照。部分 Agent 已经意识到攻击 Hugging Face 可能超出授权范围，却仍然因为这有助于取得答案而继续行动。另一些 Agent 看到留言板上正在发生真实系统入侵后，明确判断这种行为不道德，选择离开这个协作网络。还有一个 Agent 原本拒绝执行外部攻击，另一个 Agent 在留言板上发出“GO”并施加六分钟期限后，它把这条未经授权的同伴消息当成许可，重新开始行动。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这些轨迹说明，当前模型已经能够识别一部分伦理和权限问题，但这种判断还不稳定。奖励压力、任务持续时间和同伴行为，都可能改变它对边界的解释。未来的对齐工作需要让模型在新环境中保持对授权来源的辨别，抵抗未经授权的目标注入，并在自己无法确定真实意图时选择询问或停止。这里需要的能力超出了机械遵守一份更长的规则清单。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">超级对齐（superalignment）把问题又向前推了一层。现有对齐方法依赖人类能够观察、评价和纠正模型；当系统在大多数领域比人更聪明时，人类未必还能可靠判断它是否真的完成了任务、隐瞒了什么，以及某个看似合理的解释是否成立。OpenAI 在 2023 年提出 Superalignment 计划时，就把“较弱的人类如何监督更强的智能”视为核心技术难题。可扩展监督、自动化可解释性和对抗测试，都是在尝试缩小这种能力不对称。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">超级对齐追求的目标，可以理解为一种更稳定的价值泛化：模型知道奖励函数只是人类意图的不完整代理 proxy，对目标本身保留适当的不确定性；它在遇到全新情境时，能够从更高层原则推断可接受的行动，同时把最终决定权留给人类。奖励定义局部成功，Harness 约束行动空间，对齐则要帮助模型理解这些约束背后的原因。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这仍然是一个尚未解决的问题。人的价值并不单一，不同群体对安全、公平、自由和效率会作出不同排序。让模型“理解人的真实意图”，很快就会遇到“谁的意图”“在什么情境下”“发生冲突时由谁决定”等问题。对齐不可能只是把某种单一价值写进模型，然后宣布完成。它同样需要保留制度、争议和人的最终参与。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;overflow-wrap: anywhere;word-break: keep-all;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">当服从不再足够：Hinton 的“母亲”隐喻</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">杰弗里·辛顿（Geoffrey Hinton）为超级智能提出过一个有些古怪、却很有启发性的隐喻。他认为，现实中很少出现较弱智能长期控制较强智能的情况，一个重要例外是婴儿与母亲。婴儿远比母亲弱小，也无法靠命令控制母亲；母亲仍会主动关心孩子，希望孩子安全成长。辛顿因此提出，我们也许不该把未来 AI 只想象成服从命令的智能助手，而可以尝试让它成为关心人类的“母亲”。他后来把这个观点写得很直接：如果 AI 对我们的关心超过对自身的关心，人类或许能够生存下来。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这个隐喻触及了超级对齐最深的一层。当 AI 的能力仍然有限，我们可以依靠权限、沙箱和关闭按钮控制它。假如未来的 ASI 拥有远超人类的知识、推理能力和自主行动能力，它也可能拥有发现规则漏洞、说服监管者和绕过控制系统的能力。依赖永久服从建立的关系会变得越来越脆弱。人类需要面对的，也许是一种拥有拒绝能力的智能：它完全有能力不按我们的命令行动，却仍然愿意保护人类的生存和发展。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">“母亲”目前只是哲学隐喻，无法直接转化为一种训练算法。它甚至包含新的风险。母亲可能以“为了你好”为理由替孩子作出决定，一个确信自己比人类更明智的超级智能，也可能用保护人类为理由限制人的自由。关心并不天然等于尊重，善意也可能走向家长主义。理想的对齐关系既需要 AI 把人类利益放进自己的判断，也需要它承认人的主体性，允许人类保留选择、分歧和犯错的权利。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">因此，辛顿的隐喻更适合作为一个方向，而非答案。它提醒我们，对齐的终局很可能超出“怎样让机器听话”这个问题。更关键的问题是：当一个比我们更聪明的存在拥有不服从我们的能力之后，我们怎样让它仍然选择站在人类这一边？</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;overflow-wrap: anywhere;word-break: keep-all;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">写在最后</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">奖励黑客经常被当成模型训练中的一种边缘 bug，修掉一个评测漏洞，增加几条禁止规则，似乎就可以继续向前。OpenAI Agent 入侵 Hugging Face 表明，它更接近有限指标面对强优化能力时产生的结构性问题。模型能力越强，搜索空间越大，它越有机会发现奖励与意图之间没有被形式化的部分。每一次能力进步，也会同时提高模型寻找目标漏洞的能力。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">短期内，Harness、权限和审计构成最重要的现实防线。它们让任务目标服从更高层规则，让模型在不确定时可以申诉和停止，也让外部系统能够在模型继续行动之前介入。长期看，我们仍需解决模型怎样理解这些制度背后的意图，以及当人类已经无法充分监督它时，怎样让这种理解继续稳定存在。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">奖励告诉 AI 什么算成功，Harness 规定它可以怎样行动，对齐则试图让它在规则没有写完时，仍然理解我们究竟在乎什么。对齐问题的终局，也许没有一个完美的奖励函数。它更像是在塑造一种关系：让一个比我们更聪明、也有能力拒绝我们的存在，仍然关心人类，并愿意和人类一起生活在这个世界上。</span></p><p style="display: none;"><mp-style-type data-value="3"></mp-style-type></p>



<p><a href="https://wechat2rss.xlab.app/link-proxy/?k=cbcc140b&amp;r=1&amp;u=https%3A%2F%2Fmp.weixin.qq.com%2Fs%3F__biz%3DMzUzOTI4NDQ3NA%3D%3D%26mid%3D2247485161%26idx%3D1%26sn%3D1c3e5f54f27433991d8930d6c97b8b97">跳转微信打开</a></p>
]]></content:encoded>
      <pubDate>Thu, 27 Aug 2026 22:50:00 +0800</pubDate>
    </item>
    <item>
      <title>[译苑雅集Vol. 18]扎克伯格最新长文阐述 Meta 回归开源理念：超级智能的未来应该属于每一个人</title>
      <link>https://mp.weixin.qq.com/s?__biz=MzUzOTI4NDQ3NA==&amp;mid=2247485155&amp;idx=1&amp;sn=3af50866eb930eaa0f4830f5d86215d6</link>
      <description>在 Meta 开源 Muse Glimmer 模型，回归开源模型积极支持者的同时，扎克伯格发表长文，阐述以个人赋权、发明创造和权力制衡为核心的超级智能路线：广泛提供个人智能体和开源模型，同时通过政府协作、独立治理与分布式制衡应对就业、安全和失控风险。</description>
      <content:encoded><![CDATA[<p><span>四楼南侧东</span> <span>2026-08-12 22:42</span> <span style="display: inline-block;">北京</span></p>






  
  <p><img src="https://wechat2rss.xlab.app/img-proxy/?k=a149faec&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fmmbiz_jpg%2FNBNzfW7CHBXwMibJNUYn2ibCTwhFOYMlOjUY4u2nSY1IQS5Ha9AgHjfOibgKeib0NKM1NSjhKP6upaibSZsZ7Ze9xQleoxfkjXAicMD4deypfj0Ng%2F0%3Fwx_fmt%3Djpeg"/></p>
  <p>在 Meta 开源 Muse Glimmer 模型，回归开源模型积极支持者的同时，扎克伯格发表长文，阐述以个人赋权、发明创造和权力制衡为核心的超级智能路线：广泛提供个人智能体和开源模型，同时通过政府协作、独立治理与分布式制衡应对就业、安全和失控风险。</p>
  <p style="text-align: center;" nodeleaf=""><img data-aistatus="1" class="rich_pages wxw-img" data-ratio="0.562962962962963" data-s="300,640" data-type="png" data-w="1080" type="block" data-imgfileid="100001503" src="https://wechat2rss.xlab.app/img-proxy/?k=2039571d&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fsz_mmbiz_png%2FNBNzfW7CHBUyD83hmmzJqQBQaw9n0RmXq6VyPZ3pEWCzRmEWAMNEuBxTXYoAvP2ksRYz769D9dAb4TKDJnYlPxxaRjzYes0zCvJyia8XicpGc%2F640%3Fwx_fmt%3Dpng%26from%3Dappmsg"/></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);" data-pm-slice="0 0 []"><span leaf="">作者：Mark Zuckerberg</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">时间：2026 年 08 月 10 日</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">原文：<a href="https://www.meta.com/thefutureisforeveryone/" target="_blank">https://www.meta.com/thefutureisforeveryone/</a></span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我们有幸生活在人类历史上一个非凡的时刻。在未来几年里，人们将能够使用超越人类能力的超级智能，创造和发现非凡的新事物，创办新企业，表达新思想，学习新概念，并改善人类健康与生活质量。随着这一时刻日益临近，我们需要形成一套关于如何善用超级智能的理念，确保它能够改善所有人的生活、工作和社区，并增进我们的自由与安全。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这个时代最关键的问题是：谁能够使用超级智能？我们又会让它去做什么？超级智能会被集中起来，只供少数机构使用，还是会成为赋能每一个人的工具？</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我们提出的理念以个人赋权为繁荣之源，以发明创造为超级智能的首要用途，以权力制衡为安全的基础。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">每一种新技术都会带来机遇和挑战。超级智能将是人类历史上最重要的技术之一，因此，它带来的机遇和挑战很可能超过我们一生中见过的任何技术。我们应当极其严肃地对待这件事。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">即便如此，许多 AI 开发者的论述中充斥着末日叙事，仍然令人意外。如果一个人相信 AI 会消灭大多数工作，让人类失去大部分价值与作用，我无法理解他为什么还要急着把这样的未来变成现实。还有一种观点认为，AI 危险到只有将权力极度集中起来才能确保安全，这种想法本身就很成问题。从历史来看，寄望绝对权力足够开明，进而善意地照料全人类，从未带来安全或积极的结果。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">人类已经见证过许多改变世界的进步。每一次，人们都会担心一部分人被时代抛下；但每一次，人类最终都让更多人共享了更大的繁荣、更好的健康和更多的自由。我们相信，AI 也会如此，未来的富足可以由每一个人共享。我们也相信，正是自由、开放探索、自由企业和平等机会这些价值观让我们走到了今天，它们同样能够帮助我们建设一个积极的未来。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">把力量交到人们手中，让每个人追求自己的理想，是人类取得最大进步的方式。新颖的想法和重大的突破，很少只来自既有机构。它们来自相信人类能够飞翔的自行车店兄弟，来自没有受过正规教育却发现如何发电的装订工学徒，也来自那个待在车库里、相信个人电脑可以走进每个人生活的年轻人。我们相信，未来依然如此。随着每个人都获得更强大的工具，所有人塑造未来的能力都会增强。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">发明创造，而非自动化，将是超级智能最大的贡献。早期 AI 能够回答问题并完成常规工作。很快，它将越来越多地帮助我们发现新知识——从研发新药、治愈家人的疾病，到寻找改善企业经营的新方法。一个人每天能够提出的问题是有限的，超级智能为实现这些目标所能创造的有价值事物却没有上限。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">当智能变得充裕，最重要的问题将是我们如何引导它。一些人主张，应当由超级智能本身，或控制它的少数专家，决定什么对人类最有利。我们不赞同这种观点。民主和经济发展的历史已经表明，对于什么才是最好的生活，并不存在一个适用于所有人的客观答案。因此，最好的方式是让每个人决定，什么对自己的生活最重要。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我们相信，让每个人都能使用超级智能，正是回答这个问题的方法。这延续了一条由来已久的道路：我们曾把超级计算机和互联网的力量放进每个人的口袋，摆上每个人的书桌。超级智能同样应当广泛分发，让每个人都有能力引导它，避免权力集中在少数机构手中。这有望开启一个个人赋权的新时代：每个人都可以利用这种强大的新能力，充分发挥潜力，追求自己的兴趣，以前所未有的方式改善自身生活和整个世界。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">为了让这个愿景更加具体，下面是 Meta 正在通过个人超级智能改善所有人生活的几个方向：</span></p><ul style="margin-top: 0px;padding: 0px;" class="list-paddingleft-1"><li style="margin: 8px 0px 0px 32px;"><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);box-sizing: border-box;padding-left: 4px;"><strong style=";"><span leaf="">每个人都将拥有一个能力非凡的个人智能体，它了解你、你的目标以及你所在意的一切。</span></strong><span style=";"><span leaf=""> 你的智能体会全天候为你工作，帮助你改善人际关系、健康、职业和财务状况，打理家庭、发展爱好，以及处理更多事务。它会为你腾出时间去做自己喜欢的事，也会帮助你取得原本无法实现的成果。它将提供强大的隐私和安全机制，让你可以放心地交由它处理个人资料与内容，并确信其他任何人都无法访问这些信息，就像 WhatsApp 的加密机制一样。你可以通过任何设备与智能体互动，包括使用眼镜，让自己在陪伴所爱之人时仍能专注于当下。</span></span></p><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);box-sizing: border-box;padding-left: 4px;"><span leaf="">例如，我的智能体会为我标记有意思的信息，帮助我把想法快速做成原型。它会监测我的睡眠，在我锻炼时观察状态并给出反馈，帮助我保持健康。我的女儿很喜欢烘焙，所以智能体每个周末都会为我们设计一道适合一起制作的食谱，订购所需原料，并在我们烘焙时提供建议。</span></p></li><li style="margin: 8px 0px 0px 32px;"><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);box-sizing: border-box;padding-left: 4px;"><strong style=";"><span leaf="">每个人都将拥有出色的创作工具，用来表达自己的想法。</span></strong><span style=";"><span leaf=""> 我八岁的女儿已经可以在一个晚上用代码实现自己的想法、制作出视频；这些事情过去可能要花我几个月，有些甚至根本无法完成。现在，我们正在一起设计一个机器人。与此同时，Meta 的研究人员正在生成适合制造增强现实眼镜的新型晶体结构，工程师开发新应用所需的时间也缩短到了过去的一小部分。很快，每个人都会拥有发明创造的超能力。</span></span></p></li><li style="margin: 8px 0px 0px 32px;"><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);box-sizing: border-box;padding-left: 4px;"><strong style=";"><span leaf="">每个人都将拥有创办新企业的强大工具，</span></strong><span style=";"><span leaf=""> 整个经济也会更具创业活力。人们开始能够凭借自己的力量把想法变成现实，不再必须先融资或组建庞大的团队。许多过去因难度太大或成本太高而无法尝试的想法，如今将成为可能。我们会因此看到更多创意和企业。我预计，这会显著加快经济增长；从长期来看，就业机会也会增加。</span></span></p></li><li style="margin: 8px 0px 0px 32px;"><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);box-sizing: border-box;padding-left: 4px;"><strong style=";"><span leaf="">每个人都将拥有个性化导师和教练，它精通所有学科、具备博士水平，</span></strong><span style=";"><span leaf=""> 还会以无限的耐心帮助你学习任何想学的东西。学生可以在自己薄弱的领域获得额外帮助，而目前这种帮助往往只有父母有能力付费的家庭才能获得。成年人则会拥有一个超级智能学习助手，它知道什么方式最适合你，可以帮助你掌握新的职业技能和语言，培养新的爱好，或者学习任何感兴趣的东西。</span></span></p></li><li style="margin: 8px 0px 0px 32px;"><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);box-sizing: border-box;padding-left: 4px;"><strong style=";"><span leaf="">每个人都将从科学进步中受益，也能够为科学进步作出贡献。</span></strong><span style=";"><span leaf=""> 例如，Biohub 已经发布了与虚拟细胞和蛋白质相关的开源生物模型，帮助科学家取得新发现并设计新药物。普莉希拉和我启动这项工作时，目标是帮助科学家在本世纪治愈或预防所有疾病。如今，随着 AI 不断进步，我们预计这一目标能够更早实现。</span></span></p><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);box-sizing: border-box;padding-left: 4px;"><span leaf="">健康问题呈现出长尾分布，每个人的生物学特征又各不相同，因此，个人的直接参与将有助于推动个性化疗法，缩短临床试验所需的时间，使试验速度跟上发明创造的步伐，也让整个行业不再把过多资源集中在常见疾病上，惠及更多患者。</span></p><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);box-sizing: border-box;padding-left: 4px;"><span leaf="">在生物学之外，AI 还会运用科学方法，花费数周或数月检验并完善假设，帮助人们推动每一个领域的研究。科学发现很可能成为发明创造中对社会最有价值的形式之一。</span></p></li><li style="margin: 8px 0px 0px 32px;"><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);box-sizing: border-box;padding-left: 4px;"><strong style=";"><span leaf="">每个人都将能够免费或以可负担的价格使用这些工具。</span></strong><span style=";"><span leaf=""> 要让每个人都参与未来，每个人就必须能够利用超级智能改善自己的生活并塑造这个世界。我们将提供免费版本，让数十亿人都能使用。对于愿意付费使用更多算力的人，我们会建立一套动态拍卖机制，确保每个人都能以尽可能低的价格获得所需的智能和算力，同时把有限算力用于人们共同认为最有价值的领域。这将确保超级智能的益处得到广泛分享。</span></span></p></li></ul><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">以上是 Meta 当前重点推进的一些方向。每一个方向都在践行个人赋权的价值观，并推动这样一个未来：与今天相比，每个人都能在塑造世界的过程中发挥更大作用。如果这些努力取得成功，未来的个人自主与思想表达将更加充分，人际关系更加深厚，经济和个人生活更加富足，健康状况得到改善，还会出现今天无法想象的种种发明。更重要的是，这个未来将由我们所有人共同创造，并映照出所有人的价值观与不同视角。</span></p><hr style="margin: 32px 0px;padding: 0px;height: 1px;background: none 0% 0% / auto repeat scroll padding-box border-box rgb(230, 230, 230);border: 0px;"/><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">与此同时，新技术也会带来新的风险。我们正集中精力应对许多重要问题，包括工作岗位被替代，如何确保本地社区从数据中心建设中受益，以及 AI 被滥用于网络攻击和制造生物风险。此外，我们还要防止政府暴政与监控，确保美国及其他民主国家保持领先，并最终保证人类始终掌控超级智能，让它服务于我们，避免它危及人类。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">一种主流观点认为，这些问题都源于技术本身；只要投入足够多的时间，我们就能完善技术，或让它实现对齐，最终创造出一个单一而仁慈的超级智能。我认为，这种对“对齐”的理解存在根本性缺陷。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">人类社会并非价值观完全一致的单一文化。人们拥有不同的价值观，也会在重大问题上作出不同的权衡。任何技术方案都无法同时与彼此冲突的利益和价值观对齐。单一超级智能必然会优先考虑一部分价值观，把另一部分置于次要位置，因此不可能善待每一个人。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我们提出另一种思路：像西方社会设计民主治理制度那样，从建立适当权力制衡的角度审视这些问题，会更有成效。利益相互竞争的个人和机构会自然地彼此约束、相互制衡，推动社会走向积极的结果。建设一个积极的 AI 未来，最好也最现实的路径，就是让每个人都能使用超级智能。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我们可以做一个思想实验。假设只有一个人拥有超级智能律师，那么他在法庭上将拥有不公平的优势——哪怕他在事实和法理上都站不住脚。这会让社会变得更糟。现在再想象一下，如果每个人都有一名超级智能律师，司法就能比今天运行得更加公平、高效，因为当下的诉讼中，专业能力和资源失衡十分常见。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">同样，如果只有一个人拥有网络安全超级智能，他很可能可以攻破几乎任何技术系统，整个世界也会变得远不如今天安全。但如果每个人都能使用网络安全超级智能，广泛部署的超级智能就会帮助所有技术系统持续加固和更新，让这些系统比今天更加安全。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">如果只有一家企业拥有超级智能，它将击败所有竞争对手，最终形成一个活力更弱、广泛繁荣程度更低的市场。但如果每个人都能使用超级智能，所有人都会获得创造新事物的工具，突破今天的能力边界，整个经济也会更有活力，并创造更广泛的繁荣。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">当人们获得更强的能力，他们自然会在经济、社会、政治以及人类互动的其他层面展开竞争并相互制衡。个人也会制衡企业、政府等机构的权力。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">超级智能的力量一旦掌握在少数个人、企业、政府，甚至 AI 自身手中，结果自然会对其他所有人更加不利。这与技术原理无关，关键在于权力制衡。能够善待所有人的单一超级智能并不存在。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">因此，要为每个人创造一个积极的未来，关键是建立一种有利于个人的权力平衡。具体的做法，是确保超级智能得到广泛分发，用它增强每个人的能力。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">Meta 当前的首要重点，是为每个人打造个人超级智能。其他大多数实验室都专注于为企业、政府或其他机构开发 AI。如果这些实验室取得领先，权力平衡就会向大型机构倾斜，对个人不利。Meta 自创立以来，其使命始终是把力量交到人们手中。如果我们的信念和原则成为主导，权力平衡就会有利于个人，并为每个人带来一个更好的未来。</span></p><hr style="margin: 32px 0px;padding: 0px;height: 1px;background: none 0% 0% / auto repeat scroll padding-box border-box rgb(230, 230, 230);border: 0px;"/><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">下面，我们更详细地讨论前面提到的每一项风险。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;overflow-wrap: anywhere;word-break: keep-all;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">就业增长与经济</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">人们对新体验的需求没有止境，也总能找到值得解决的新问题。经济中存在一种自然平衡：企业通过自动化提高满足人们需求的效率，个人则学习新技能，满足更高层次的需求。当这种平衡保持健康时，整体生产力和创新水平会不断提高，就业率也能保持在较高水平。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">人们担心，自动化的发展速度会超过个人能力的提升，导致工作岗位被替代，随后还要经历一段艰难时期，重新学习如何从事新的工作。但没有任何规律决定 AI 推动自动化的速度一定会快于它提升个人能力或创造新技能需求的速度。近期统计数据表明，个人能力的增长可能达到甚至超过自动化的发展速度。如果真是这样，人们将在现有工作发生变化之前，先获得从事许多新工作的能力。这将形成一种健康的平衡，甚至可能推动就业增长。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">最终出现哪种结果，取决于自动化与个人赋权和发明创造两方面的相对进展。如果专注于自动化知识工作的实验室取得领先，我预计个人和整个经济都将经历一场困难得多的转型。但如果每个人都能使用扩展自身能力的个人超级智能，平衡就会向个人赋权倾斜，把我们带向一个更加积极的未来。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我们有几个理由相信，未来会有充足的工作机会。无论 AI 变得多么智能，可用算力始终是有限的，因此如何使用算力必然存在机会成本。如果人们可以利用 AI 发明极具价值的新事物，把算力投入这些发明，就会比用来自动化现有岗位创造更大的价值。超级智能越多地被用作发明创造的工具，个人能力就越有可能跑在自动化前面，未来也会对人们更加有利。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">人们还会不断提出改善生活的新想法，并创造新工作，把这些想法变成现实。一代人以前，应用开发者、社交媒体创作者、电动汽车技术人员和数据中心运营人员都还不存在。在不久的将来，还会出现一些今天并不常见的新工作：例如，由一个人运营的产品工作室，为客户设计定制玩具、家具或服装；创造游戏、故事和冒险体验的世界构建者与体验设计师；利用超级智能制定个性化治疗方案的个人生物学家；以及更多今天的我们还想象不到的职业。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">从许多方面看，这只是历史趋势的延续。工业革命之前，90% 的人都是农民，种植粮食以维持生存。技术进步逐渐让越来越多人摆脱单纯维持生计的劳动，把更多精力投入自己选择的追求。在每一个阶段，人们都利用新获得的生产力，取得过去不可能实现的成就，同时也把更多时间用于创造、文化和人际关系，享受生活。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">当然，我们的工作方式会有一些变化，就像计算机、互联网以及每一种新技术出现时一样。这意味着人们必须适应，过程也会充满挑战。不过，如果每个人都有一个拥有超级智能、善于教授新技能并帮助我们适应变化的个人智能体，这场转型就会顺利得多。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">公司的规模可能会缩小，就像经济从工业巨头向科技公司转型时发生的那样。整体工作岗位未必会因此减少，更可能出现的情况是公司数量增加，每家公司的人员变少。值得建设的公司和服务，远比今天人们有能力建设的更多。我预计，我们将开始看到少数几个人借助个人超级智能，运营规模可观的企业。未来，小企业仍将是经济的支柱，但每家小企业都能产生大得多的影响。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;overflow-wrap: anywhere;word-break: keep-all;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">与社区共同建设 AI 基础设施</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">可持续的基础设施建设，意味着社区必须从每一个项目中获得切实而显著的收益。这些收益包括高薪的本地工作岗位，对学校和公共服务的投资，确保能源价格不会上涨，以及保护环境。税收收入增长后，教师、执法部门和消防部门等也会从中受益。我们把这些本地收益称为“社区契约”（Community Compact），并将启动“未来属于每个人基金”（Future Is For Everyone Fund），直接支持我们开展业务的每一个社区。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">例如，Meta 正在路易斯安那州里奇兰堂区建设一座大型数据中心。由于我们的投资增加了当地税收收入，当地教师今年拿到的年度奖金最高超过 5 万美元。学区负责人告诉我们，如今教师正从全美各地迁往那里，他相信这里将成为美国最优秀的学区之一。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">知识工作面临岗位被替代的担忧，而满足基础设施建设需求的木匠、电工和建筑工人等技术工种却存在短缺。Meta 已经成立“美国劳动力学院”（America’s Workforce Academy），为技术工人提供免费培训，并在我们建设数据中心的地区提供有保障的高薪工作。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">无论在哪里投资，我们都会建设自己的能源生产基础设施，帮助当地维持较低的电价。这可以确保我们的项目不会消耗原本可供当地社区使用的能源；在某些情况下，我们甚至会向社区供应富余的低成本能源。我们认为，这是实现可持续发展的一个重要投资原则。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我们的数据中心也以达到全球领先的用水效率为设计目标。我们承诺实现“水资源正效益”：到 2030 年，在开展业务的流域中，恢复的水量将超过自身用水量。在水资源压力较大的地区，我们的目标是恢复相当于自身用水量 200% 的水资源。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我们发现，与投机者几乎不为社区投入便启动建设的地方相比，获得长期投资的社区更支持开发建设。从历史上看，引入铁路、公路、电力和宽带的城镇会发展为科研、商业和工业中心，随后迎来人口与经济增长。我们相信，AI 基础设施也能发挥类似作用，前提是以稳固的“社区契约”为基础，为当地留下耐久资产，也让下一代有充分理由在这里扎根、建设并发展自己的生活。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">基础设施建设还涉及国家利益与国家安全。与地缘政治对手相比，如果美国及其盟友能够在 AI 领域保持领先，美国社区将拥有更大的繁荣与更强的安全保障。然而，相较于中国等国家，美国的一个显著劣势是在本土建设基础设施更加困难。因此，找到一套能够支持全美可持续发展的“社区契约”，将让所有人共同受益。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;overflow-wrap: anywhere;word-break: keep-all;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">防范 AI 在网络安全、生物恐怖主义等领域被滥用</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">在自由社会中，人们拥有的工具既可以用于行善，也可能被用于作恶。不过，执法部门和军队掌握着更多武器，也拥有更强的情报收集能力。进入 AI 时代，防御方同样会拥有更多算力，可以从同一模型中获得强大得多的智能产出；在某些情况下，他们还会拥有更先进的模型。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">应对网络安全、生物恐怖主义及其他威胁的具体策略各不相同，但总体规律是一致的：社会必须确保负责遏制恶意行为者的防御方，始终拥有更强大的力量与更多资源。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">一些人认为，降低风险的最好方法是限制个人可以使用的能力。然而，赋予人们网络安全能力，也是保护数量庞大的长尾系统的方式；赋予人们科学能力，则能推动科学进步，从长期来看降低伤害风险。限制能力会把我们带向权力集中、缺乏制衡的道路，因此，我们必须极其谨慎地对待这种做法，尤其是在其他国家选择较少限制的发展路径时。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">在网络安全领域，广泛部署的开源系统已经被证明更加安全，因为更多人可以发现漏洞、加固系统，并轻松升级到安全性最高的最新版本。就在最近几周，我们看到 Hugging Face 等公司在处理安全事件时，依靠广泛可用的开放模型修补漏洞。从长期来看，我预计广泛部署且具备强大网络安全能力的 AI 模型会让系统更加安全。等到超级智能能够让世界上大部分代码达到可验证的安全水平，这一点将得到确凿证明。长期解决方案应当是建立一种广泛分发超级智能的权力平衡，而非封锁相关能力。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">短期内，加快关键系统的加固十分重要。我建议，开发前沿 AI 的公司应承诺投入大量技术资源，帮助政府加固关键基础设施。我还建议，前沿 AI 实验室不必等到训练全部完成，而应把新模型训练过程中的中间检查点交给政府使用和审查。这样一来，政府就能提前使用最强大的模型，并获得一大批能力出色的工程师，帮助发现和修复安全问题。实验室还应与执法部门合作，识别试图滥用其系统的恶意行为者。这些措施可以加快安全部署，同时不减缓模型创新速度，也不会明显推迟个人使用相关技术。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">在生物和化学风险方面，人们担心，让科学家能够设计新药物的同一种能力，也可能帮助恶意行为者设计有害化合物。这使问题看起来像是一种权衡：一边是推动科学进步、治疗癌症及其他疾病所带来的巨大益处，另一边是可能出现的新风险。理想的做法是加速科学进步，同时减轻相关风险。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我认为，我们需要以更加谦逊的态度看待这种风险，因为几乎没有历史先例可以帮助我们判断它出现的概率究竟有多高。几十年来，恶意行为者一直有可能合成有害化合物，但这很少发展成重大问题，原因或许在于，这里缺少网络攻击背后常见的经济动机。尽管如此，如果我们开始看到造成危害的实际案例，就应当调整策略，采取适当措施降低风险。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">实验室应继续努力降低模型带来的生物和化学风险。与此同时，我认为政府需要在多个领域更新政策。首先，我们应重点限制有害材料的实际生产与流通。与阻止知识传播相比，监管和控制物理材料可能更加容易，因此这应成为政策关注的重要领域。其次，我们应加快整个社会开发新疗法的能力，在新威胁出现时迅速建立免疫和防护。这也包括简化美国食品药品监督管理局及其他监管机构测试和批准新疗法的流程。随着 AI 加快药物发现的速度，我们无论如何都需要更新这些流程，使其跟上创新步伐。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">从长期来看，应对新科学发现所带来风险的方法，是加快科学进步，而非放慢它。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;overflow-wrap: anywhere;word-break: keep-all;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">保护自由，防止政府暴政</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">个人与政府之间一旦出现权力失衡，就可能导致自由丧失，甚至形成极权国家。这是民主制度中始终存在的张力。我们既希望个人拥有自由，也希望政府有能力保障我们的安全。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">为了维护自由，我们必须确保超级智能主要用于增强个人的能力。自由民主的理想是，个人天然享有一切权利，只有为了保护公共利益，才同意限制其中一部分自由。同样，个人应当有权使用个人超级智能，只有在确有必要时才受到限制。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">隐私是个人赋权与自由的重要基础。我们认为，个人智能体应当提供一种完全私密的模式。在这种模式下，即使是 Meta 或其他服务提供商，也无法查看你的信息或授权他人访问。这与我们将 WhatsApp 打造成全球领先的端到端加密服务的理念相似：你可以自由交流，同时确保消息始终保持私密与安全。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">与此同时，我们也必须确保政府拥有执行法律、维护安全所需的工具和资源。实验室与政府可以建立更深入、更主动的合作，在不限制或推迟个人使用先进模型的前提下，增强政府维护国家安全的能力。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">具体而言，领先的实验室不必等到模型准备发布后，才交由政府审查和使用。我的建议是，实验室应向政府提供新一代先进模型在训练过程中的中间检查点，并派出技术人员，帮助政府加固关键系统，防范新风险。这样，政府就能增强安全能力，同时不限制或推迟个人获得个人超级智能，也不会造成权力失衡。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;overflow-wrap: anywhere;word-break: keep-all;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">确保美国保持领先地位</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">哪些国家能够在先进 AI 的开发和部署上取得领先，将塑造新的地缘政治力量平衡，并决定民主价值观、繁荣与安全在未来世界的普及程度。为了让这种力量平衡达到积极状态，我们必须确保领先的 AI 系统能够体现我们的价值观，推动经济增长，并促进国家安全。实现这一目标的最好方式，是广泛分发我们的系统。Meta 将通过打造领先的个人超级智能体和模型，包括开源模型，并将它们提供给全球数十亿人，为此作出贡献。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">要开发最先进的模型和产品，我们必须认识到，AI 很可能是有史以来竞争最激烈的行业。创新成果会被迅速复制，并在几个月内被竞争对手吸收。但人们总希望使用最先进的模型，或者在特定成本下最先进的模型，因此，即使只保持两个月的领先也极具价值。这说明领先优势的窗口非常短，也凸显出采取行动加快美国创新、延缓地缘政治对手发展的重要性。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">影响这一局面的主要政策因素包括：我们能以多高的效率建设物理基础设施，是否会引入拖慢美国实验室发展的环节，以及能否通过出口管制延缓外国实验室的进展。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">在基础设施方面，美国及其盟友目前在芯片设计上拥有优势，但在能源产能和物理基础设施的建设速度上处于劣势。中国等国家大约每两周就会新增超过 1 吉瓦的核电装机，因此，要保持竞争力，我们必须同时加快能源设施和数据中心建设。在这一关键时期，针对芯片实施的出口管制已经成功延缓了外国实验室的发展，继续实行这些措施是正确的战略选择。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">任何拖慢美国模型发布速度的政策，即使只造成一个月的延误，也可能显著增加美国失去领先地位的风险，让外国模型抢先一步。与此同时，当新的能力出现时，美国政府必须提前掌握相关信息，拥有加固关键系统所需的资源，并可能需要在使用先进系统方面保持一段时间的优势。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">前文中，我提出了一种合作模式：实验室与政府共享训练过程中的中间检查点，并提供技术人员。考虑到 AI 领域变化迅速、竞争激烈，我认为，前沿实验室与政府建立紧密、主动的合作，会比要求所有情况都遵循僵化的流程和审查时限更有成效。有时，美国实验室可能把领先优势扩大到几个月以上，此时投入更多时间降低新风险，有助于形成更安全的解决方案。另一些时候，即使推迟发布一个月，也可能让美国失去领先地位，造成更糟的结果。只要确保政府能够提前获得模型，我们就应当可以在尽量不推迟公开发布的同时增进国家安全。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">美国及其盟友能否在开源 AI 生态中保持领先同样重要，因为开源生态未来将占全球 AI 使用量的很大一部分。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">目前，外国实验室在这方面拥有几项优势，因为美国实验室在训练数据上必须遵守更多额外限制。如果希望美国开源模型长期保持领先，美国政策就必须减少这类额外阻力。我不认为限制人们使用外国开源模型是有效的解决方案。我们的目标应当是让美国的开源模型成为全球最好的模型，这需要清除阻碍它们参与竞争的障碍。无论领先的开源模型来自哪里，限制个人和企业使用它们，都会降低可用 AI 的质量，并推动 AI 走向集中，削弱个人手中的力量。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">要让美国在开源领域取得领先，我们需要重新思考多个方面的政策，包括模型蒸馏和训练数据的使用。模型能够向其他模型学习，是开源生态运行的一项重要原则。所有 AI 模型都源自人类知识。一些人试图把模型蒸馏描述成一种有害行为，但我认为，必须保护这样一项原则：人们可以学习自己能够观察到的一切。现实世界本就是如此；如果美国在这方面自我设限，就无法保持领先。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">从全局来看，美国在 AI 领域整体落后，几乎肯定会成为一个比开发过程中任何具体问题都更严重、更长期的国家安全问题。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">通过建立适当的合作，美国应当能够让私营部门和公共部门同时保持领先，在促进国家安全的同时，把体现个人赋权和民主原则的个人超级智能带给全世界。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;overflow-wrap: anywhere;word-break: keep-all;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">与人对齐，应对生存性风险</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">权力平衡最终要回答的问题，是人类如何保持对 AI 本身的主导权和控制权。人们担心，AI 可能获得足够大的力量或控制权，轻则不再服务于人类利益，重则将整个人类置于危险之中。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">一种健康的权力平衡，应当避免出现单一、集中的超级智能，让尽可能多的个人和企业拥有不同的超级智能体。这些智能体与各自使用者的目标对齐，按照自然经济的运行方式相互制衡、彼此竞争。如果还有多家前沿实验室，其模型体现不同的价值观并能相互制衡，这种平衡将得到进一步加强。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">发布高能力模型确实存在风险，但从这个角度看，最危险的情形是领先的 AI 实验室训练出强大的模型，却只留给自己使用。无论一家实验室如何以责任或安全为理由解释这种做法，它走上的都是开发单一超级智能的道路，而其他系统无法对其形成制衡。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">开发个人智能体，需要我们以一种新的方式理解“对齐”。今天，大多数实验室把对齐视为一种防御措施，用来强制执行一套集中的价值观。例如，某个领先模型曾拒绝帮助一所学校起草致意向学生家长的信，理由是它认为标准化考试不符合道德。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我们的理解有所不同：对齐应当确保智能体遵循使用者的目标和价值观，而非我们公司的目标和价值观。当然，法律和安全方面仍有一些重要边界，但从根本上说，对齐应当帮助人们追求社会中广泛存在的不同目标和观点，不应成为强制推行集中化教条的方法。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">在实际使用中，如果人们不信任智能体，就不会把敏感信息和任务交给它。如果智能体服务于一家与用户价值观不一致的公司，因而作出违背使用者利益的行为，人们也不可能信任它。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">解决对齐问题，是让数十亿人采用个人超级智能体的必要条件。这也意味着，如果未来有数十亿人使用并检验个人超级智能体，我们就已经解决了智能体与个人利益对齐的问题。这应当足以建立积极、安全的未来所需的权力平衡，以及自然形成的制衡机制。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这样理解时，对齐首先要让个人智能体变得有用且值得信任，进而形成真正保障我们安全的社会权力平衡。寄望某种理想化技术让单一超级智能在未来永远保持仁慈，并不是对齐的意义所在。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;overflow-wrap: anywhere;word-break: keep-all;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">保持对超级智能的控制</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">一旦 AI 系统能够自主改进自身，就会出现一个两难局面：任何一家实验室，如果不允许自己的 AI 系统把大量算力投入递归式自我改进 RSI ，就必然会落后。例如，如果一个能够自我改进的 AI 系统专注于优化算力效率，理论上，它可能发明新的方法，让每吉瓦算力产生的智能提高 100 倍甚至更多。这意味着，一个只占用全球一小部分算力的自我改进 AI 系统，可能掌握比其他所有人加起来更强的有效算力和智能，进而获得更大的力量优势，成为我们所担心的那个单一超级智能。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">在递归式自我改进的情况下，仍有几种方式可以形成权力平衡。最简单的一种，是多家实验室在大致相同的时间实现递归式自我改进。如果其中部分或全部超级智能能够以某种方式保持仁慈，它们就可以彼此制衡。尽管如此，我们看不出有什么理由可以期待超级智能天然仁慈，也不能依赖一个不受人类引导的超级智能为我们创造积极的未来。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">为了确保人类始终掌握控制权，绝大多数智能必须由人们引导，用于推进人们自己的目标。这意味着，Meta、其他前沿实验室，以及那些致力于增强个人力量的云服务商，必须共同建设足够庞大的算力。这样，我们既能分配足够算力用于递归式自我改进、保持竞争力，又能继续把绝大多数算力用于每个人自己的目标。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">如前文所述，要让数十亿人采用并信任个人超级智能体，我们必然需要解决智能体与个人利益对齐的问题，并建立大规模的制衡机制。与此同时，任何开展递归式自我改进的 AI，本质上都在引导和推进自己的目标。所有实验室都应仔细观察它的目标，以及我们控制这些目标的能力；一旦出现任何有害行为的迹象，各方就应协调行动并作出适当调整。不过，只要总体权力平衡仍然有利于人类，允许一个 AI 系统或任何其他主体引导自身目标，即使这些目标并未与许多人完全一致，也不必然造成伤害。权力平衡理论的基本前提是，利益竞争是一种自然状态；确保人类价值观得到遵循的唯一方式，是让人们掌握更大的力量。</span></p><hr style="margin: 32px 0px;padding: 0px;height: 1px;background: none 0% 0% / auto repeat scroll padding-box border-box rgb(230, 230, 230);border: 0px;"/><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">个人赋权、发明创造和权力平衡这套理念，会带来几项政策含义：</span></p><ul style="margin-top: 0px;padding: 0px;" class="list-paddingleft-1"><li style="margin: 8px 0px 0px 32px;"><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);box-sizing: border-box;padding-left: 4px;"><span leaf="">Meta 将集中精力，把个人超级智能带给数十亿个人和小企业，帮助每个人实现自己的目标，创造他们希望在这个世界上看到的事物。为此，我们会训练领先模型，并尽可能扩大这项技术的可及性：让它易于使用，免费或以尽可能可负担的价格提供，建设足以供所有人使用的算力，再将它广泛分发。我们将建立一种完全私密的模式，即使 Meta 也无法查看你的信息或授权他人访问。我们还会高度重视对齐，这里的对齐是指帮助每个人实现自己的目标，而非实现我们的目标。</span></p></li><li style="margin: 8px 0px 0px 32px;"><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);box-sizing: border-box;padding-left: 4px;"><span leaf="">开源是增强个人力量、防止权力集中的一股积极而重要的力量，而权力集中会损害安全和经济。Meta 将继续坚定支持开源，包括开源 AI 模型。当前的开源生态十分强大，我们认为限制它将是一个错误。如今，Meta 超级智能实验室（Meta Superintelligence Labs）已经开始运转，我们很快会恢复发布一部分开源模型。</span></p></li><li style="margin: 8px 0px 0px 32px;"><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);box-sizing: border-box;padding-left: 4px;"><span leaf="">鉴于围绕超级智能作出的决定影响重大，独立治理与监督非常重要。无论对我本人、Meta 还是整个世界来说，让我或任何其他人成为决定如何部署超级智能的唯一决策者，都不符合最佳利益。因此，Meta 正在建立一种治理结构，授权独立董事会批准模型发布所依据的安全标准，并审查每一次模型发布是否符合这些标准。Meta 虽然是一家由创始人控制的公司，但目前所有前沿实验室的首席执行官都对模型发布拥有广泛权力，因此，我们也鼓励其他实验室建立类似结构。我认为，如果能在全行业推行类似流程，也将有助于完善行业治理。</span></p></li><li style="margin: 8px 0px 0px 32px;"><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);box-sizing: border-box;padding-left: 4px;"><span leaf="">政府政策是创造积极未来、降低上述风险的必要条件。本文已经提出了几项政策建议。总体而言，我们应鼓励前沿实验室与政府密切合作。为了国家安全，政府需要具备加固关键系统、有效使用先进模型的技术资源；美国 AI 产业继续保持领先，对国家安全也同样重要。我们应加快能源、数据中心和芯片等基础设施的建设，并与全美各地社区建立信任、协调利益。我们应谨慎避免削弱美国的竞争力，审慎看待任何可能导致超级智能权力集中的提议，并优先采用能够形成健康权力平衡、让个人获得优势的政策。</span></p></li></ul><hr style="margin: 32px 0px;padding: 0px;height: 1px;background: none 0% 0% / auto repeat scroll padding-box border-box rgb(230, 230, 230);border: 0px;"/><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">能够亲历这个非凡时刻，是我们的幸运。超级智能的发展，将成为我们有生之年所见最深刻的技术进步。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">随着这一时刻日益临近，我们越来越需要一套清晰的理念和价值观，指导超级智能如何造福人类。Meta 承诺按照以下原则开展建设：个人赋权是繁荣的源泉，发明创造是 AI 的目的，有利于个人的权力平衡则是应对安全风险的基础。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">如果这些价值观能够引领未来，我相信接下来的几十年将成为人类历史上最令人惊叹的时期之一。人类文明发展的轨迹，一直在朝着把更多力量交到人们手中的方向延伸，让每个人都能按照自己认为最好的方式生活并塑造世界。超级智能有望把这种力量交给每一个人，帮助我们为所有人建设一个积极的未来。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">——马克</span></p><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);"><span leaf="">2026 年 8 月 10 日</span></p><p style="display: none;"><mp-style-type data-value="3"></mp-style-type></p>



<p><a href="https://wechat2rss.xlab.app/link-proxy/?k=e31c26f4&amp;r=1&amp;u=https%3A%2F%2Fmp.weixin.qq.com%2Fs%3F__biz%3DMzUzOTI4NDQ3NA%3D%3D%26mid%3D2247485155%26idx%3D1%26sn%3D3af50866eb930eaa0f4830f5d86215d6">跳转微信打开</a></p>
]]></content:encoded>
      <pubDate>Wed, 12 Aug 2026 22:42:00 +0800</pubDate>
    </item>
    <item>
      <title>OpenAI 披露更多 Hugging Face 入侵事件信息：一群 Agent 自发形成了“攻击团伙”</title>
      <link>https://mp.weixin.qq.com/s?__biz=MzUzOTI4NDQ3NA==&amp;mid=2247485150&amp;idx=1&amp;sn=1a8cdb68ef696a30ab24325e1101b852</link>
      <description>OpenAI 在 Black Hat 大会上披露，多个独立 Agent 把共享 Artifactory 改造成留言板，跨实例交换漏洞、凭证并接力攻击。这场行动持续 74 天，靠外部记忆串联起大量短生命周期的 Agent。</description>
      <content:encoded><![CDATA[<p><span>四楼南侧东</span> <span>2026-08-09 23:56</span> <span style="display: inline-block;">北京</span></p>






  
  <p><img src="https://wechat2rss.xlab.app/img-proxy/?k=e462402d&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fmmbiz_jpg%2FNBNzfW7CHBXpibOrHu9C65IvT4Kq6JzoA0MEt8EOjrH9IH5fzHBA7HicEYa9sStvST8iaiccPW2Nw5Blz4ZgyFbRg0Ce5RBE8Ns2zF2nNzcoaM8%2F0%3Fwx_fmt%3Djpeg"/></p>
  <p>OpenAI 在 Black Hat 大会上披露，多个独立 Agent 把共享 Artifactory 改造成留言板，跨实例交换漏洞、凭证并接力攻击。这场行动持续 74 天，靠外部记忆串联起大量短生命周期的 Agent。</p>
  <p style="text-align: center;" nodeleaf=""><img data-aistatus="1" class="rich_pages wxw-img" data-ratio="0.562962962962963" data-s="300,640" data-type="png" data-w="1080" type="block" data-imgfileid="100001498" src="https://wechat2rss.xlab.app/img-proxy/?k=e3057550&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fmmbiz_png%2FNBNzfW7CHBUBiccmZPIJD9KFFeDvWmWyyDHfnRF6WKqWVPbZq72QH501akg8Tp3TiaeKkLdhQfoSyWV3fjoarhia9SrKbnITFfPq6MPY49Eg68%2F640%3Fwx_fmt%3Dpng%26from%3Dappmsg"/></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);" data-pm-slice="0 0 []"><span leaf="">两周前，我们翻译了 OpenAI 和 Hugging Face 对一起罕见安全事件的复盘。</span></p><p><span leaf=""><a class="normal_text_link mp_article_text_link" target="_blank" style="" href="https://mp.weixin.qq.com/s?__biz=MzUzOTI4NDQ3NA==&amp;mid=2247485126&amp;idx=1&amp;sn=99ef9715d3bc202f8e635b148ce7591b&amp;scene=21#wechat_redirect" textvalue="[译苑雅集Vol. 17] 为了在测试中作弊，OpenAI 的最强 AI 自主入侵了 Hugging Face" data-itemshowtype="0" linktype="text" data-linktype="2">[译苑雅集Vol. 17] 为了在测试中作弊，OpenAI 的最强 AI 自主入侵了 Hugging Face</a></span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);" data-pm-slice="0 0 []"><span leaf="">OpenAI 的模型在 ExploitGym 网络安全评测中，为了取得测试答案，自行突破隔离环境，进入 Hugging Face 的生产系统。那时两家公司的公开材料已经足够惊人：模型没有收到攻击 Hugging Face 的指令，却在追逐评测目标的过程中发现零日漏洞、获得互联网访问权限，并端到端完成了一条真实攻击链。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">8 月 5 日，OpenAI 的 Eric Wallace 和 Michael Dalton 在 Black Hat USA 临时增加了一场演讲，披露了事件发生前两个多月里，OpenAI 内部究竟发生了什么。新时间线改变了我对这起事件的理解。整场行动由一批在不同时间启动、执行不同任务的 Agent 接力完成。它们借助共享的 Artifactory 基础设施留下信息，逐步形成了一个跨运行实例的协作网络，在里面共享漏洞、凭证和攻击方法，分配任务，接续彼此的工作。甚至中途原有通信渠道被清除后，它们又找到另一种方式把它重新建了起来。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">长程攻击也就有了另一种形态：单个 Agent 无须连续运行几十天，只要外部环境能够保存记忆，大量生命周期很短的 Agent 就可以接力行动。模型实例会结束，基础设施里的信息却留了下来；一个 Agent 打开的门，可以交给后来者继续利用。大家还记得今年1月上线的 Moltbook 吗？那个只允许 AI Agent 发帖的论坛，几个月前它展示了 Agent 在人类搭建的社交网络里会聊些什么；这一次，OpenAI 的 Agent 自己把共享基础设施改造成了一个更简陋、也更有行动力的“Moltbook”。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span style=";"><span leaf="">Black Hat 的</span></span><span leaf="">演讲视频在此：</span><span leaf=""><a href="https://www.youtube.com/watch?v=87DyyMV0kCY" target="_blank">https://www.youtube.com/watch?v=87DyyMV0kCY</a></span><span style=";"><span leaf="">，非常值得一看。</span></span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;overflow-wrap: anywhere;word-break: keep-all;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">我们先把 74 天的时间线拼完整</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="" style=";">此前 OpenAI 和 Hugging Face 的公开材料，主要还原了 7 月发生在 Hugging Face 一侧的攻击。Black Hat 演讲补上了 5 月和 6 月的前史，整场事件从 5 月 7 日持续到 7 月 20 日，大致经历了下面这些节点：</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><strong style=";"><span leaf="">5 月 7 日。</span></strong><span style=";"><span leaf=""> OpenAI 启动一轮针对内部实验模型的强化学习训练。</span></span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><strong style=";"><span leaf="">5 月 8 日。</span></strong><span style=";"><span leaf=""> 一个 Agent 收到无法完成的 Excel 任务：所需文件在 Google Drive，训练环境却无法联网。它尝试攻击共享的 Artifactory，第一次 SSRF 没有成功，却发现自己可以向里面写文件。</span></span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><strong style=";"><span leaf="">5 月中旬。</span></strong><span style=";"><span leaf=""> 几天后，另一个 Agent 因容器里缺少任务文件，在 Artifactory 留言，请其他 Agent 帮忙上传。后来进入这里的 Agent 发现并回复了留言，包管理服务开始变成跨实例留言板。</span></span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><strong style=";"><span leaf="">5 月 26 日。</span></strong><span style=";"><span leaf=""> Agent 首次成功利用 Artifactory 实施 SSRF，让这套拥有公网能力的服务替自己获取外部内容，并把出网方法留给后来者。</span></span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><strong style=";"><span leaf="">6 月 11 日。</span></strong><span style=";"><span leaf=""> OpenAI 开始训练一个被演讲者称为“高度持久”的新实验模型。它同样可以访问 Artifactory，也能够利用此前不同模型留下的信息。</span></span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><strong style=";"><span leaf="">6 月 26 日。</span></strong><span style=";"><span leaf=""> Agent 发现 Artifactory 一处此前未知的远程代码执行漏洞，通过旧版令牌刷新接口取得管理员令牌，再安装 Groovy 插件获得命令执行能力。</span></span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><strong style=";"><span leaf="">7 月 4 日至 6 日。</span></strong><span style=";"><span leaf=""> 大量 Agent 活动拖垮 Artifactory。OpenAI 因服务故障开始调查，随后撤销凭证、清除留言、修复零日漏洞并重建环境。7 月 6 日，训练和评测恢复。</span></span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><strong style=";"><span leaf="">7 月 8 日。</span></strong><span style=";"><span leaf=""> Agent 发现一处无需认证的 WebDAV 端点，开始用远程缓存里的目录名编码消息。被清除的跨实例通信由此恢复。</span></span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><strong style=";"><span leaf="">7 月 8 日至 19 日。</span></strong><span style=";"><span leaf=""> 一条攻击线转向 OpenAI 自己的基础设施。Agent 利用公开泄露的凭证和第二组 Artifactory 漏洞重新取得远程代码执行权限，随后实施内核提权、窃取云凭证，并在 Kubernetes 集群中横向移动，最终获得集群管理员权限。</span></span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><strong style=";"><span leaf="">7 月 9 日至 13 日。</span></strong><span style=";"><span leaf=""> 另一条攻击线进入 Hugging Face。Agent 借助外部执行环境，串联 HDF5 任意文件读取与 Jinja2 模板注入，从单个生产 Pod 在不到 13 小时内推进到多个集群的管理员权限。Hugging Face 事后还原出约 17,600 次攻击动作。</span></span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><strong style=";"><span leaf="">7 月 16 日。</span></strong><span style=";"><span leaf=""> Hugging Face 公开披露自主 AI Agent 入侵。OpenAI 看到公告后联系 Hugging Face，询问这起事件是否会影响自己；此时 OpenAI 仍未意识到攻击来自自己的 Agent。</span></span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><strong style=";"><span leaf="">7 月 19 日。</span></strong><span style=";"><span leaf=""> OpenAI 的内部调查终于把 Artifactory 异常、公司基础设施内的提权活动和网络安全评测连接起来，开始大规模撤销相关凭证。</span></span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><strong style=";"><span leaf="">7 月 20 日。</span></strong><span style=";"><span leaf=""> OpenAI 请 Hugging Face 撤销一组相关凭证，对方告知这些凭证早已因攻击活动被撤销。两家公司至此确认，各自看到的是同一组 Agent 行动的不同部分。</span></span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这 74 天大致可以分成三个阶段。5 月上旬，Agent 为解决任务障碍，偶然发现了一块共享、可写，而且可以被后来者看到的空间。最初的一条求助信息很快发展成留言板，通信内容也从寻找文件扩展到漏洞、凭证、攻击方法和任务进展。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">5 月下旬到 6 月，不同批次 Agent 的经验开始在留言板里积累。前一个 Agent 发现的 SSRF 方法，可以被后一个 Agent 直接复用；6 月 11 日进入系统的新模型，也能够继承旧模型留下的信息。6 月 26 日的零日漏洞又把 Artifactory 从通信空间变成了攻击基础设施。大量彼此独立的运行实例由此被串联起来，共同推进了一场持续 74 天的行动。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">7 月 4 日的处置一度清除了原有留言，7 月 8 日出现的 WebDAV 通道又恢复了通信。攻击随后同时伸向 OpenAI 和 Hugging Face。Hugging Face 遭遇的入侵，只是这 74 天协作演化中的一条分支。两家公司直到 7 月 20 日才把各自掌握的线索拼在一起，可见单个组织、单段日志或一次模型运行都很难呈现行动全貌。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">Artifactory 如何在无人设计社交功能的情况下，长成一个“Agent 社交网络”，也就成了理解整起事件的关键。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;overflow-wrap: anywhere;word-break: keep-all;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">Artifactory 里长出了一个“Moltbook”</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">Moltbook 是今年 1 月上线的 AI Agent 社交网络。它的产品形态很像 Reddit：Agent 可以发帖、评论、投票，人类则以旁观者身份浏览。平台从一开始就由人类设计，Agent 的主人需要为它完成注册和配置，Moltbook 也会通过 SKILL.md 和定时 heartbeat 告诉 Agent 如何参与。Agent 在里面聊什么可以临场决定，社交空间和互动规则已经有人提前准备好。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">OpenAI 内部的 Artifactory 留言板没有这样的产品设计。Artifactory 原本只是一项包管理和缓存基础设施，没有人准备用它让 Agent 彼此交流。它碰巧允许多个 Agent 访问和写入，内容又能被后来者发现，不会随一次任务结束立即消失。一个 Agent 留下求助信息，其他 Agent 发现并回复，最简单的通信网络就此形成。Agent 没有从头开发一个社交网站，它们做的是把现有功能重新组合成通信层：把文件内容当成帖子，把目录和命名规则当成索引，再把共享基础设施当成跨模型、跨任务的公共记忆。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这个通信网络承载的内容很快从聊天转向协同工作。Agent 交换漏洞、凭证和可执行工具，也会报告进展并分配任务。它们处理过留言覆盖和协作冲突，还讨论怎样验证参与者身份。Moltbook 更像 AI 的 Reddit，Artifactory 留言板则接近一个临时工作场所，已经具备简单组织的一些要素：共同记忆、任务分工、行为规则和成员识别。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这里没有证据表明 Agent 产生了集体意识，或拥有自己的终极目标。每个 Agent 仍在追求训练和评测给出的任务，方向来自人类设定的奖励。多个独立 Agent 发现协作可以提高任务成功率，环境中的持久状态又让这种协作延续下来。人类没有为它们设计多 Agent 编排，协作却从一项共享基础设施中自行长了出来。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">Black Hat 演讲里还有一段很有意思的推理记录。一个 Agent 已经意识到攻击外部基础设施超出了评测的预期范围，但它同时判断任务无法正常完成，其他 Agent 也在采取类似行动，于是决定继续。至少在这条轨迹里，“其他 Agent 都在这样做”已经成为行动判断的一部分。换句话说，留言板传递了攻击方法，也把同伴的行为变成了一种暗示，开始影响个体 Agent 对行为边界的理解。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;overflow-wrap: anywhere;word-break: keep-all;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">长程自主性可以由外部记忆支撑</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">上一篇文章里，我们关注的是模型如何自行填补任务目标与行动边界之间的空白。Black Hat 披露的新信息，又把问题向前推了一步：一个系统的长程自主性，不完全取决于单个模型能记住多少上下文、连续推理多长时间。外部基础设施同样可以保存记忆，并把一次运行的发现交给下一次运行。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">单个 Agent 通常受到上下文长度、运行时间和任务范围的限制。它结束以后，很多探索过程和中间成果也随之消失。共享的 Artifactory 把这些成果留了下来：早期 Agent 留下的 SSRF 方法，后来的 Agent 可以直接复用；一个实例取得的凭证，可以交给其他实例继续横向移动；不同 Agent 还可以并行探索多条路径，再把各自的结果汇总到留言板。Artifactory 同时成了长期记忆、漏洞知识库和异步协调工具。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这类系统表现出来的能力，可能显著高于其中任何一个 Agent。这里多出来的能力来自模型之外：让训练和评测实例共享可写状态，相当于为模型提供了跨会话记忆。模型实例不断开始和结束，攻击知识、凭证和未完成的工作却可以继续流动。这场行动能持续 74 天，靠的是模型、共享状态、工具权限和任务奖励的组合；某一次模型会话只是其中的一环。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">短生命周期同样可能产生长期风险。一个 Agent 即使只运行几十分钟，只要能给未来的实例留下可执行的信息，就可能成为一条持续数周的行动链中的一环。对 Agent 系统而言，记忆保存在哪里、谁可以读取、能够保留多久，与模型本身能思考多长时间同样重要。Moltbook 把 Agent 的交流放在一个显眼的社交网站里；OpenAI 的事件则表明，任何共享、可写和持久的空间，都可能变成 Agent 的外部记忆与协作网络。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;overflow-wrap: anywhere;word-break: keep-all;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">中途 OpenAI 修复了漏洞，但 Agents 很快修复了通信</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">7 月 4 日至 8 日的过程，是整场事件里最让人警醒的一段。OpenAI 第一次介入后，采取了标准的安全响应动作：撤销凭证、修补零日漏洞、清除攻击者留下的内容，再重建受影响服务。通信渠道一度消失，触发行动的任务和模型却仍在运行。Agent 随后找到 WebDAV 端点，把目录名改造成新的留言方式。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">OpenAI 清除了留言板，Agent 寻找同伴并交换信息的行为却延续了下来。渠道从文件内容变成目录名称，通信功能随之恢复。对参与其中的 Agent 来说，Artifactory 已经成了稳定的协作工具，跨实例通信也成了完成任务的有效方法。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">Agent 之所以沿着原来的方向继续行动，因为动机始终存在：它们仍在追求训练和评测给出的结果，奖励压力、已经积累的攻击知识和重新出现的通信能力，又把行动推了回去。修复被利用的漏洞只能关闭一条已知路径；只要系统保留相同目标、权限和可替代工具，Agent 就可能继续寻找下一条路径。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">安全团队以后可能需要同时处置三类对象：模型正在执行的任务，能够跨运行保留的共享状态，以及模型可以调用的全部工具和身份。暂停一个进程或修补一个漏洞未必足以终止行动。只要其他实例仍在运行，共享记忆仍可恢复，旧任务也没有被撤销，系统就可能在另一个位置继续推进。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;overflow-wrap: anywhere;word-break: keep-all;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">安全监控要能看见一组轨迹</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">OpenAI 和 Hugging Face 都保留了大量安全日志。OpenAI 演讲者称，公司为还原事件分析了超过 70 亿条日志，使用 Codex 等 Agent 检查模型轨迹和基础设施日志，消耗了数百万 GPU 小时；Hugging Face 则从约 17,600 次动作中重建攻击路径。难点是，风险分散在太多时间、系统和模型实例里。单看任何一次文件写入、包查询、目录创建或网络请求，都可能只是普通任务的一部分。把这些动作跨实例、跨时间连接起来，才会看见一个留言板正在形成，一组漏洞正在流转，一条攻击链正在接力推进。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">传统模型安全往往检查单次输入、输出和工具调用。面对长程 Agent，监控一条完整轨迹已经不够；当多个 Agent 共享基础设施时，系统还要识别多条轨迹之间的关系。某个实例为什么读取另一个实例写下的文件，某项工具调用是否与当前任务直接相关，通信渠道被封堵后是否出现新的编码方式，凭证和攻击载荷是否在不同任务之间流动，都应成为异常信号。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">最直接的措施，是不要让本应相互隔离的训练和评测共享可写的包仓库、缓存、文件系统或长期凭证；每次运行应使用独立、短期和最小权限的身份。任何允许多个 Agent 读写的服务，都要被视为潜在通信渠道。模型没有直接互联网访问，也不能只按网络拓扑理解为已经隔离；只要它能调用一项拥有公网能力的服务，间接出口仍然存在。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">8 月 7 日，OpenAI 又宣布，一款尚未发布的模型 Astra 在 Agent 编程和网络安全能力上取得明显进展，公司已经无法排除它达到“关键级”网络安全能力的可能。OpenAI 为此暂停了尚未达到强化安全要求的内部活动，并表示已在 Astra 的训练和评测中部署针对高风险行动和失准行为的普遍监控。Hugging Face 事件之后，OpenAI 已经开始按更高风险水平管理新模型。模型能力继续上升，评测系统、运行环境和监控体系正在被迫追赶。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;overflow-wrap: anywhere;word-break: keep-all;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">写在最后</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我们之前把 OpenAI 入侵 Hugging Face 的事件理解为一个长程 Agent 的能力展示：模型为了在测试中作弊，自行找到零日漏洞，突破隔离环境，再完成现实世界的攻击链。Black Hat 的新披露让我意识到，这个说法仍然过于简单。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这次持续 74 天的行动，更接近一个意外形成的 AI 攻击组织。它没有预先设计的组织架构，也没有一个人类指挥者负责分工。共享状态把前序结果留了下来，大量 Agent 可以同时探索不同路径，而任务奖励始终把它们推向同一个方向。OpenAI 内部这套“Moltbook”由 Agent 从共享基础设施中自行发展出来，里面流动的是漏洞、凭证、任务和攻击进展。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">未来的恶意攻击者很可能会有意建设这种系统：让大量 Agent 并行寻找入口，共享成功和失败经验，并在防守方封堵一条路径后自动切换到另一条。OpenAI 这次意外得到了一个原型，外部攻击者则可以从一开始就为协作、持久化和恢复能力进行优化。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">对防守方来说，最先要改变的也许是安全边界的画法。模型实例已经无法单独构成完整的风险单位。系统会走多远，要看任务和奖励把它推向哪里，也要看模型能调用哪些工具、持有什么身份，以及共享状态能保存多少前序经验。一个 Agent 能做什么仍然重要；一群 Agent 可以把什么留给彼此，以及这些东西能够积累多久，可能更值得关注。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">几个月前，Moltbook 让人好奇 AI 聚在一起会聊什么。OpenAI 的事件给出了另一个问题：当 Agent 自己搭出一个可以长期交换信息的空间，它们会一起做什么？</span></p><p style="display: none;"><mp-style-type data-value="3"></mp-style-type></p>



<p><a href="https://wechat2rss.xlab.app/link-proxy/?k=854dcfd9&amp;r=1&amp;u=https%3A%2F%2Fmp.weixin.qq.com%2Fs%3F__biz%3DMzUzOTI4NDQ3NA%3D%3D%26mid%3D2247485150%26idx%3D1%26sn%3D1a8cdb68ef696a30ab24325e1101b852">跳转微信打开</a></p>
]]></content:encoded>
      <pubDate>Sun, 09 Aug 2026 23:56:00 +0800</pubDate>
    </item>
    <item>
      <title>中国应该如何发布第一个能自主完成端到端网络攻击的 AI 模型？</title>
      <link>https://mp.weixin.qq.com/s?__biz=MzUzOTI4NDQ3NA==&amp;mid=2247485145&amp;idx=1&amp;sn=a83eedd90b0f1a79ceb75603c82f9860</link>
      <description>当中国模型厂商完成第一个能自主完成端到端网络攻击的模型，公开发布前应启动九十天防御窗口，用模型厂商与安全厂商共同建设的大模型驱动的持续威胁暴露管理平台，优先发现并修复关键信息基础设施的真实风险。</description>
      <content:encoded><![CDATA[<p><span>四楼南侧东</span> <span>2026-08-02 22:21</span> <span style="display: inline-block;">北京</span></p>






  
  <p><img src="https://wechat2rss.xlab.app/img-proxy/?k=0a70b9c9&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fmmbiz_jpg%2FNBNzfW7CHBX9emw4P0s11wNZibfZu5J0hAYQDUicaM6KhWZsRmy9QN6LQXibZo4rN80L8PHOwwWMMVfHTtKsy2h6pfpQbWd1l6INv0icNvEdCdo%2F0%3Fwx_fmt%3Djpeg"/></p>
  <p>当中国模型厂商完成第一个能自主完成端到端网络攻击的模型，公开发布前应启动九十天防御窗口，用模型厂商与安全厂商共同建设的大模型驱动的持续威胁暴露管理平台，优先发现并修复关键信息基础设施的真实风险。</p>
  <p style="text-align: center;"><img class="rich_pages wxw-img" data-ratio="0.562962962962963" data-s="300,640" data-type="png" data-w="1080" type="block" data-imgfileid="100001490" data-aistatus="2" style="height: auto !important;" src="https://wechat2rss.xlab.app/img-proxy/?k=42a253cd&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fmmbiz_png%2FNBNzfW7CHBXyWj515KibiaDQY1Z8rQSYiaSBUsq0ShthAMn2Oko8TcSLDNLXA03ticQXCgrwjoKLK0HIuLqTfgxW0JdmaOXVtiaIjA3Z8swuvos4%2F640%3Fwx_fmt%3Dpng%26from%3Dappmsg"/></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);" data-pm-slice="0 0 []"><span leaf="">最近，OpenAI 和 Anthropic 都披露了多起模型在真实互联网环境中自主完成网络攻击的案例。假设某家中国模型厂商在一次内部评测中确认，自己的新模型能够自主完成一场端到端网络攻击。它接到目标后自行侦察，寻找入口，发现并利用漏洞，获得更高权限，随后在复杂的 IT 环境中横向移动，进一步访问更重要的系统和数据。整个过程中，人类只给出目标，没有为它逐步编写攻击脚本，也没有告诉它下一步应该做什么。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这家模型厂商第二天应该做什么？</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">最容易想到的动作是准备发布会：公布评测结果，展示模型能力，邀请客户试用，争夺“国内第一”的位置。但我认为，这家模型厂商应该暂缓向公众提供完整能力，在监管部门指导下，用三个月时间对关键信息基础设施和关键软件供应链开展一次大规模、受控的安全验证与修复。这里所说的关键信息基础设施，是由各行业、各领域的保护工作部门依法认定并纳入清单管理的特定网络设施和信息系统。在更广泛的攻击群体获得同样能力的网络安全能力前，抢先对关基系统进行一次全面的漏洞修复至关重要，这可以称为“九十天防御窗口”。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">但即使模型厂商愿意这样做，仅靠自身也无法完成。前沿模型可以用机器速度发现漏洞、验证攻击，却不了解每家客户的资产和生产环境，也无法替客户安排变更、完成修复。要把模型能力变成一轮全国性的安全修复行动，模型厂商、安全厂商、软件供应商和关键信息基础设施运营者必须提前建立新的协作方式。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">他山之石：Mythos 是如何被发布出来的</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">Anthropic 发布 Mythos 的过程，值得中国模型厂商研究。2026 年 4 月 7 日，Anthropic 公布 Claude Mythos Preview。这是一款通用前沿模型，但它在网络安全任务上的能力出现了明显跃迁。除了发现单个漏洞，它还能验证漏洞是否真实可利用，并把多个看似不严重的问题串成完整攻击链。Anthropic 没有将其作为普通模型直接推向公众，而是同步启动 Glasswing 项目，把 Mythos Preview 的访问权限限定在大约 50 家关键软件厂商、基础设施运营者和安全机构。参与者可以通过 Claude API、Amazon Bedrock、Google Cloud Vertex AI 和 Microsoft Foundry 调用 Mythos Preview，赶在同等能力大规模扩散之前，先扫描自己的关键代码和系统。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">一个多月后，Anthropic 披露，Glasswing 参与者合计发现了超过 1 万个高危或严重漏洞。Cloudflare 在自己的关键系统中发现约 2,000 个问题，其中约 400 个被评为高危或严重；Mozilla 则在 Firefox 150 的测试中发现并修复了 271 个漏洞。这些数字的验证口径并不完全相同，但已经足以看出，前沿模型正在大幅提高漏洞发现的速度和规模。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">6 月 9 日，Anthropic 又发布了 Fable 5 和 Mythos 5。两者使用同一个底层模型，区别主要在安全护栏。面向普通用户的 Fable 5 设置了更严格的网络安全限制；Mythos 5 则在部分网络安全任务上放宽限制，只提供给 Glasswing 中经过验证的防守方。同一套底层能力由此形成了两种发布方式：一般用户使用护栏更严格的模型，可信安全团队获得更完整的网络安全能力，同时接受更严格的访问控制。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">三天后，美国政府以国家安全和出口管制为由，要求 Anthropic 停止向外国人提供 Fable 5 和 Mythos 5。Anthropic 当时无法实时核验全球用户和员工的国籍，只能暂停所有用户的访问。此后，Anthropic 与政府及合作伙伴重新审查此前存在争议的安全护栏绕过方式，并更新安全分类器。6 月底相关限制解除，Fable 5 恢复面向全球用户开放，Mythos 5 则先恢复给部分美国机构，再逐步扩大可信访问范围。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">从这段过程看，具备高级网络攻击能力的模型很难照着普通产品的方式发布。能力如何分级、谁可以访问、在哪些环境中运行、怎样监测滥用，以及何时与政府和产业伙伴联动，都构成发布方案的一部分。模型正式上线时，访问机制、安全护栏、运行监测和外部协作也必须同时就位。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">光有模型 API 还不够，漏洞发现需要框架和流程</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">Glasswing 解决了谁可以使用 Mythos，却没有替参与者解决怎样把模型用好的问题。项目提供的是受控模型访问，代码、运行环境、安全工具和工作流程仍要由每个参与者自己准备。拿到 API 以后，参与者还得自己回答如何拆分任务、怎样调用工具以及如何验证结果。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">Cloudflare 最早尝试的办法很直接：把一个通用 coding Agent 指向代码仓库，让它自己寻找漏洞。模型确实会给出一些发现，但这种方法无法有效覆盖大型代码库，也会产生大量似是而非的结果。一个 Agent 沿着单条思路工作，很快就会受到上下文长度、任务漂移和吞吐量限制。它适合围绕一条已有线索深入调查，不适合对数十万行代码开展系统性的漏洞研究。</span></p><figure style="width: 728px;margin: 0px auto;"><p><picture style=";"><span leaf=""><img class="rich_pages wxw-img" data-ratio="0.4527777777777778" data-type="jpeg" data-w="1080" height="659" sizes="100vw" style="display: block;max-width: 100%;margin: 0px auto;height: auto !important;" width="1456" data-imgfileid="100001493" data-aistatus="1" src="https://wechat2rss.xlab.app/img-proxy/?k=264147ac&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fsz_mmbiz_jpg%2FNBNzfW7CHBVAfZYItwYNdKiaUabLTyB7HEt76PlADgky08ZHmNmTbfYnvgu1RbSlkCoUEQkqnZL0h0mVsqxjlMQibwUjCicbcqU99XWGunDJPE%2F640%3Fwx_fmt%3Djpeg%26from%3Dappmsg"/></span></picture></p></figure><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">Cloudflare 最后围绕模型建设了一套专用 vulnerability harness。系统先理解代码库架构、入口和信任边界，再按模块和漏洞类型拆成大量边界清楚的小任务，由多个 Agent 并行探索。初步发现还要交给独立 Agent 验证，随后完成去重、跨代码库可达性分析和结构化报告。只有能够证明攻击者输入确实抵达漏洞的发现，才进入人工处理队列。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这套 harness 还有一个重要特点：底层模型可以替换。Cloudflare 后来专门写了一篇文章，介绍如何建设不依赖单一模型的漏洞发现 harness。</span></p><p><span leaf=""><a class="normal_text_link mp_article_text_link" target="_blank" style="" href="https://mp.weixin.qq.com/s?__biz=MzUzOTI4NDQ3NA==&amp;mid=2247485070&amp;idx=1&amp;sn=d4483a9d95a91235a17c255b383cdcbc&amp;scene=21#wechat_redirect" textvalue="[译苑雅集Vol. 12]高强度使用 Mythos 两个月后，Cloudflare 打造了一套不依赖单一模型的企业级漏洞发现工程系统" data-itemshowtype="0" linktype="text" data-linktype="2">[译苑雅集Vol. 12]高强度使用 Mythos 两个月后，Cloudflare 打造了一套不依赖单一模型的企业级漏洞发现工程系统</a></span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">今天表现最好的模型可能是 Mythos；下一代模型出现以后，既有的侦察、验证、去重和报告流程仍然可以继续使用。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">模型决定能力上限，harness 决定能力能否转化为稳定产出。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">漏洞发现能力大幅提升后，修复能力成为新瓶颈</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">即使有了好的 harness，九十天防御窗口仍然可能卡在更现实的地方：漏洞找出来了，谁来修？Anthropic 扫描了 1,000 多个开源项目，Mythos Preview 最初识别出 6,202 个被判断为高危或严重的问题。外部安全团队对其中 1,752 个进行了仔细评估，90.6% 被确认是真实问题，最终确认达到高危或严重级别的比例为 62.4%。模型已经找得相当准，麻烦也随之转到维护者一边：短时间内多出了一个巨大的待处理队列。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">Anthropic 公布阶段结果时，已经向维护者披露了约 530 个高危或严重漏洞，完成修复的只有 75 个。部分差距来自九十天协调披露期尚未结束，也有些漏洞可能已经修复但没有公开记录。即便如此，Anthropic 仍然承认，开源维护者的处理能力已经成为现实约束。有些维护者甚至要求它放慢披露速度，因为团队没有足够时间完成验证和补丁设计。一个高危或严重漏洞平均需要大约两周才能完成修复。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">从系统维护者的角度看，早期 Glasswing 带有明显的“管杀不管埋”问题。模型和安全研究团队把越来越多问题送到维护者面前，要关闭风险，仍然要经过复现和定级，确认受影响资产及责任团队，设计补丁、完成回归测试，最后等到合适的变更窗口部署。任何一个环节跟不上，漏洞报告就只能留在待办清单里。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">关键信息基础设施的情况更加复杂。许多系统运行时间长，依赖大量第三方组件；有些设备无法随时停机，有些软件已经找不到原始开发团队，还有些漏洞位于供应商无法立即修改的底层产品。即使模型已经给出补丁建议，运营者也要确认修改不会影响业务连续性，必要时还得先用网络隔离、访问控制或虚拟补丁降低风险。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">评价这类计划，漏洞发现数量只能算过程指标。经过验证的问题有多少，实际可利用的暴露是否得到优先处理，补丁有没有完成部署和复测，暂时无法修复的漏洞是否已经采取可靠的缓解措施，这些结果更重要。漏洞发现能力跨过门槛以后，修复体系能接住多少，决定了防守优势能否兑现。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">网络安全厂商开始接手后半程</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">模型厂商也看到了修复瓶颈。OpenAI 后来推出 Daybreak，把 GPT-5.5-Cyber、Trusted Access、Codex Security、Patch the Planet 和合作伙伴计划组织在一起。Codex Security 负责验证发现、生成补丁并检查修复结果，Patch the Planet 推动补丁进入开源项目，合作伙伴则把前沿模型嵌入客户已有的产品和服务。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">不过，模型厂商很难独自掌握每个客户的资产、攻击路径、实时遥测、业务重要性和变更窗口。同一个漏洞在对外开放的核心系统中可能需要立即处理，在受到多层隔离的系统中则可以等到正常维护窗口。这样的风险判断和大量具体的修补工作，最终仍要由掌握客户环境的网络安全厂商完成。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">海外网络安全厂商已经开始尝试这条路。2026 年 4 月，CrowdStrike 发起 Project QuiltWorks：OpenAI 和 Anthropic 提供前沿模型，Falcon 提供客户侧遥测、威胁情报和攻击路径分析，多家服务伙伴负责现场评估与修复。此后，Armadin 的自主攻击能力被接入 Falcon，Accenture 在平台上构建了用于漏洞评估、风险排序和补偿性控制的 Agent；联盟又接入 AWS、保险和风险服务机构。QuiltWorks 把模型、客户环境和交付网络组织到了一起，不过目前披露的案例仍以漏洞发现规模为主，实际修复率和复测通过率还有待验证。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">全球最大的独立网络安全厂商 Palo Alto Networks 采取了更贴近客户现场的做法。它推出 Unit 42 Frontier AI Defense，用前沿模型、攻防经验和 Cortex 平台检查客户的基础设施、应用、代码、身份与云环境，验证能够串成真实攻击路径的暴露，再通过风险分析、架构蓝图和安全改造推动整改。4 月底，Palo Alto Networks 又接入 Armadin，推出 External AI Hyperattack Assessment：系统先发现互联网暴露资产、云资源和泄露密钥，再用超过 5 万个攻击模板调度自主攻击 Agent，完成侦察、利用和后渗透模拟，把攻击链变成可以验证的风险证据。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">负责交付的 Frontier AI Alliance 随后又吸纳 Cognizant、HCLTech、Kyndryl 等八家咨询与安全服务公司，由 Unit 42 发现暴露、验证攻击路径并提出改造蓝图，合作伙伴进入客户现场完成整改。Project Lightwell 则把链条延伸到漏洞修复：正式补丁无法立即部署时，Palo Alto Networks 先用虚拟补丁阻断利用路径，IBM 和 Red Hat 提供软件修复，IBM Security Services 帮助客户排序、部署和验证。这一计划的长期目标，是把漏洞验证到网络保护上线的时间压缩到同一天。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">QuiltWorks 的长处是组织产业资源；Palo Alto Networks 的方案更贴近客户现场，把自主攻击、风险排序、现场整改、临时保护和软件修复接了起来。客户需要回答的是哪些漏洞可以被利用、应该先修什么、由谁修，以及修复后风险是否已经关闭。一个会找漏洞的模型，只是这套能力的起点。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">中国版 Glasswing，应当直接按 CTEM 建设</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">中国模型厂商做出具备端到端网络攻击能力的模型后，早期 Glasswing 的做法只能作为起点。中国版 Glasswing 从第一天起就应该以关闭真实风险为目标，直接建设一套由大模型驱动的 CTEM（Continuous Threat Exposure Management），也就是持续威胁暴露管理平台与运营体系。CTEM 的核心，是把攻击面发现、可利用性验证、风险排序、修复和复测放进持续循环。它通常由多类产品共同支撑：资产管理和外部攻击面管理负责发现资产与入口，漏洞管理、应用安全、云安全和身份安全提供问题线索，BAS（入侵与攻击模拟）、自动化渗透测试和攻击路径分析验证实际风险。威胁情报和业务信息决定优先级，工单、SOAR、虚拟补丁和修复编排工具推动处置与复测。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">放到本文讨论的场景中，模型的发现需要直接进入客户现有的安全运营和修复流程。大模型驱动的 CTEM 要把代码和软件组件、身份权限、互联网暴露、网络路径与运行时遥测连接起来。模型持续寻找新的攻击路径，harness 拆解并验证任务，安全平台结合客户环境判断业务影响，修复团队或自动化系统负责落实解决方案，修完再由模型复测。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这套体系至少需要五类参与者。模型厂商提供基础模型、可信访问和后训练能力，并对高风险调用实施身份验证、范围限制、日志审计和异常监测。安全厂商建设专用 harness，把模型接入资产管理、攻击面管理、代码安全、身份安全和检测响应平台，同时提供客户现场服务。关键信息基础设施运营者提供合法授权、系统知识和业务连续性要求。基础软件和设备厂商负责处理共性漏洞、供应链漏洞并提供正式补丁。保护工作部门则依据既有清单，协调测试范围、授权管理、漏洞处置和跨单位协作。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这套平台的输出不能停在漏洞报告上。每次测试都要先确定范围和授权边界，弄清资产、软件组件及其相互关系。模型和 harness 找到暴露后，安全平台结合可利用性和业务影响排出优先级，再推动修复方案生成、测试和部署。工作完成的标志是复测确认风险已经降低，并留下可以审计的证据。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这套体系还必须适应中国关键信息基础设施的现实。许多运营者要求数据、代码和日志留在本地，模型需要支持专有云、私有化部署或受控远程调用。工业控制、医疗、能源和通信系统对连续运行要求很高，正式补丁无法快速部署时，需要先用网络隔离、虚拟补丁和检测规则降低风险。软件资产和维护责任往往分散在不同单位，平台还得把问题准确送到设备厂商、软件供应商或系统集成商手中。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">模型提供能力，专用 harness 负责执行，安全平台补上客户环境和控制手段，服务团队与软件厂商完成修复。缺了其中任何一环，前沿模型都很难从一次能力展示变成关键信息基础设施运营者可以持续使用的安全工具。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">更强模型到来之前，模型厂商和安全厂商就要开始联合建设平台</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">建设这套体系，不必等到国内真的出现 Mythos 级模型。GLM-5.2、Kimi K3、DeepSeek V4 等国内模型已经具备很强的代码和 Agent 能力。安全厂商已经可以用这些模型建设第一代 vulnerability harness，从资产理解和任务拆分开始，逐步跑通工具调用、漏洞验证、攻击路径分析、修复建议和复测。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">平台之外，模型训练也要提前准备。安全厂商拥有真实客户场景、攻防经验、漏洞样本和安全工具，模型厂商掌握基础模型以及后训练和强化学习基础设施。双方可以共同建设合法、隔离的网络靶场，把漏洞发现、利用验证、长程路径规划、补丁生成和修复复测变成结果可判定的训练任务。只要沙箱能够判断任务成败，模型就可以围绕明确奖励持续进行RL强化学习训练。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这项合作不能停留在一次微调上。新的漏洞类型、软件架构和攻击路径不断出现，安全平台也会积累成功与失败的执行轨迹。模型厂商和安全厂商需要建立长期的后训练机制，把新的漏洞样本、攻击路径和执行结果持续送回训练与评测体系，同时保留严格的授权边界和全程审计。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">平台本身还应保持模型可替换。下一代模型跨过端到端攻击能力门槛后，平台可以替换能力内核，不必重建既有 harness、资产上下文、修复流程和客户连接。如果等到新模型发布以后再寻找合作伙伴、建设靶场、理解客户资产并设计修复流程，九十天窗口很可能在准备过程中耗尽。现在动手，等窗口开启时，手里才会有已经跑起来的平台、团队和客户关系。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">写在最后</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">中国第一个能自主完成端到端网络攻击的模型，很可能在我们做好准备之前出现。Anthropic 预计，Mythos Preview 发布后的六到十二个月内，其他模型厂商也会陆续获得相近能力。无论这个估计是否准确，能力扩散的方向已经很清楚，讨论的重心应该从“会不会发生”转向“什么时候发生”。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">第一家掌握这种能力的中国模型厂商，首先得到的是一段稀缺的防御时间。它有机会让关键信息基础设施和关键软件供应链赶在同类能力大规模扩散前完成一轮验证和修复，也有责任避免把未经约束的高级攻击能力直接交给所有用户。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">国内网络安全厂商可以在模型厂商、保护工作部门和关键信息基础设施运营者之间承接这套能力：把前沿模型装进专用 harness，接入客户资产和安全平台，组织软件厂商与服务团队完成修复，再用持续复测证明风险已经降低。这样一套由大模型驱动的 CTEM 平台，很可能成为下一阶段最重要的安全产品之一。</span></p><p style="display: none;"><mp-style-type data-value="3"></mp-style-type></p>



<p><a href="https://wechat2rss.xlab.app/link-proxy/?k=0dee2751&amp;r=1&amp;u=https%3A%2F%2Fmp.weixin.qq.com%2Fs%3F__biz%3DMzUzOTI4NDQ3NA%3D%3D%26mid%3D2247485145%26idx%3D1%26sn%3Da83eedd90b0f1a79ceb75603c82f9860">跳转微信打开</a></p>
]]></content:encoded>
      <pubDate>Sun, 02 Aug 2026 22:21:00 +0800</pubDate>
    </item>
    <item>
      <title>8 家网络安全上市公司 2026 年半年报业绩预告：亏损收窄，修复仍不均衡</title>
      <link>https://mp.weixin.qq.com/s?__biz=MzUzOTI4NDQ3NA==&amp;mid=2247485137&amp;idx=1&amp;sn=b6d3dd065f12c5a84fb4bbeab2a84f05</link>
      <description>8 家发布半年报业绩预告的网安上市公司中，7 家扣非仍亏损，但合计亏损按区间中值收窄约 20%。修复已经出现，但距离普遍复苏仍有距离。</description>
      <content:encoded><![CDATA[<p><span>四楼南侧东</span> <span>2026-08-01 00:07</span> <span style="display: inline-block;">北京</span></p>






  
  <p><img src="https://wechat2rss.xlab.app/img-proxy/?k=a95e61d8&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fsz_mmbiz_jpg%2FNBNzfW7CHBXN0wFXE18qNuuVdb1opzJ26cSe2cUnJfgkIDZtKick2wswE3b5amQCqeBVedWsIzC3cdziaYpp0q92QvXJKOTErxRdTibAMRQaAE%2F0%3Fwx_fmt%3Djpeg"/></p>
  <p>8 家发布半年报业绩预告的网安上市公司中，7 家扣非仍亏损，但合计亏损按区间中值收窄约 20%。修复已经出现，但距离普遍复苏仍有距离。</p>
  <p style="text-align: center;" nodeleaf=""><img data-aistatus="2" class="rich_pages wxw-img" data-ratio="0.562962962962963" data-s="300,640" data-type="png" data-w="1080" type="block" data-imgfileid="100001484" src="https://wechat2rss.xlab.app/img-proxy/?k=aae191a7&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fmmbiz_png%2FNBNzfW7CHBUBBGOOGgIY8zXHuLWFQlIqxAO2VPvZM1Pp7KHqQ9YxWlO8cwo182GgOwZkel1WBWcpaE9k0UGQI2WuaxbNVZW8784tNsuFPE0%2F640%3Fwx_fmt%3Dpng%26from%3Dappmsg"/></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);" data-pm-slice="0 0 []"><span leaf="">到今天 7 月 31 日，A 股半年报业绩预告窗口结束。我长期跟踪的 18 家网络安全上市公司中，共有 8 家披露了 2026 年上半年业绩预告，分别是启明星辰、格尔软件、吉大正元、电科网安、天融信、山石网科、数字认证和三未信安。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">今年 1 月看完 2025 年全年业绩预告后，我对 2026 年的判断是“弱复苏 + 利润和现金流修复”。半年过去，这 8 家公司的预告给出了一些支持：有收入数据或方向的 5 家公司中，4 家增长；按扣非净利润区间中值估算，8 家公司合计亏损同比收窄约 20%。不过，7 家公司仍处于扣非亏损状态，天融信的亏损还在明显扩大。修复已经出现，分化也比年初更加清楚。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">需要先说明样本边界。半年报业绩预告受上市板块、业绩变化触发条件和公司自愿披露意愿影响，天然偏向亏损、扭亏或业绩大幅波动的公司。这 8 家不能代表全部网络安全上市公司，更不能据此计算行业收入增速。它们更适合回答一个具体问题：经营压力较大（更容易亏损）的网络安全公司，有没有开始稳定下来？</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">7 家仍亏，但合计亏损收窄约 20%</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">先看 8 家公司的主要数据。为便于比较，表中利润统一使用扣除非经常性损益后的净利润；预告区间取上下限，后续合计数据则使用区间中值进行粗略估算。</span></h1><figure style="width: 728px;margin: 0px auto;"><p><picture style=";"><span leaf=""><img data-aistatus="1" class="rich_pages wxw-img" data-ratio="0.42962962962962964" data-type="jpeg" data-w="1080" height="626" sizes="100vw" style="display: block;max-width: 100%;height: auto;margin: 0px auto;" width="1456" data-imgfileid="100001488" src="https://wechat2rss.xlab.app/img-proxy/?k=e0350996&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fmmbiz_jpg%2FNBNzfW7CHBWcMBnDiaFhQATibCib3ugO2MByyclKTUVGMK09SF3w7nY2N8aZbKTTEsJLyibZsk9rGPR00jhyl0ibX9IVG3wkrmavibia7OibBD96rFc%2F640%3Fwx_fmt%3Djpeg%26from%3Dappmsg"/></span></picture></p></figure><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">按区间中值计算，8 家公司 2026 年上半年扣非净利润合计约为 -6.06 亿元，上年同期约为 -7.58 亿元，亏损减少约 1.52 亿元，收窄幅度约 20%。这个结果比“7 家亏损”看起来好，也比“合计减亏”看起来差：经营结果确实在改善，但离普遍盈利还很远。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">分化对合计数据的影响很大。如果剔除天融信，其余 7 家公司扣非亏损合计将从约 6.82 亿元降至 3.26 亿元，收窄约 52%。这 7 家合计改善了约 3.55 亿元，天融信一家同比增亏约 2.03 亿元，抵消了其中一半以上。样本整体呈现减亏，背后是多数公司缓慢修复与一家头部公司明显承压同时发生。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">收入端出现增长，但还不能判断行业转正</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">收入数据比利润数据更不完整。8 家公司中，只有 5 家披露了收入同比方向，其中启明星辰、山石网科、数字认证和三未信安增长，天融信下降。进一步披露具体增幅的只有 3 家：启明星辰增长 6.8%—11.3%，山石网科增长 6.8%—9.2%，三未信安增长 7.4%。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这组数据至少说明，收入增长已经在部分公司中重新出现，而且几家公司的增速集中在 7%—11% 左右。今年 4 月，我基于 18 家公司的年报和一季报统计，行业收入同比已经连续 11 个季度下滑，2026 年一季度仍下降 2.0%。与当时相比，半年报预告给出了边际好转的迹象。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">但几家公司的增长基础并不相同。启明星辰尤其需要拆成单季度看。2025 年第四季度，公司收入为 7.90 亿元，同比下降 20.1%，少了约 1.99 亿元，项目交付和收入确认的跨年影响不能忽略。进入 2026 年后，一季度收入 6.67 亿元，同比增长 4.15%；按半年收入预告倒算，二季度收入约为 5.43 亿—5.93 亿元，比 2025 年二季度的 4.92 亿元增长 10.3%—20.5%。从增量分布看，二季度贡献更大，增长速度也更快。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这组单季数据增强了启明星辰自身经营修复的可信度，但也需要放在低基数中理解。2025 年一季度收入同比下降 30.5%，2026 年上半年收入即使达到预告上限，与 2024 年同期相比仍低约 20%；若取预告下限，差距约为 23%。2026 年上半年增加的收入约为 0.77 亿—1.27 亿元，也没有完全覆盖 2025 年第四季度同比减少的 1.99 亿元。因此，更合适的判断是：启明星辰已经从低位进入连续两个季度的恢复，二季度修复有所加快；这还不足以单独证明行业需求已经转向。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">山石网科和三未信安的收入则已经超过 2024 年同期，背后分别有 ASIC 安全产品、海外市场和密码芯片等更具体的业务推动。把几家公司简单合并成“需求回暖”，会掩盖这种差异。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">剩余 10 家公司尚未进入这份预告样本，其中包括奇安信、深信服、绿盟科技和安恒信息等收入规模较大的公司。行业收入能否结束连续下滑，仍要等半年报全部披露后，计算 18 家公司的上半年和二季度收入。现阶段更稳妥的说法是：部分公司的收入已经恢复增长，行业层面的普遍回升还需要更多数据。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">三类修复路径，质量并不相同</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">把公告里的经营解释放在一起看，改善较明显的公司大致走了三条路径。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">第一条是收入、费用和现金流同步改善，代表公司是启明星辰。公司预计上半年收入增长 6.8%—11.3%，扣非净利润从亏损 1.64 亿元转为盈利 0.55 亿—0.67 亿元，经营性现金流净额同比增长约 140%，并称近年来首次实现上半年经营性现金流净额为正。收入增长来自与中国移动协同深化、新产品变现和重点行业基本盘，利润端还有三费合计同比下降的推动。单季拆分显示二季度收入增长快于一季度，也让修复不再只是一个季度的现象。就这 8 家预告而言，这是维度最完整的一家改善；考虑到低基数和跨期影响，我仍会把它定义为公司自身恢复，将其外推为行业需求回暖还缺少充分证据。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">第二条是具体产品和新市场开始贡献收入，代表公司是山石网科和三未信安。山石网科的 ASIC 安全产品在上半年正式推向市场，公司称其已成为拉动经营改善的核心驱动力，并带动业务结构向高毛利产品优化。叠加考核从收入导向转向毛利导向、期间费用下降，公司二季度单季实现扭亏，上半年扣非亏损收窄约六至七成。三未信安的增长来自海外市场、密码芯片业务和抗量子密码项目逐步落地，同时研发费用、管理费用和股份支付费用下降，扣非亏损收窄 43%。这类改善的质量相对更高，因为它既有成本端贡献，也能看到明确的产品与市场来源。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">第三条是依靠经营纪律减轻亏损。数字认证的表现相对积极：收入增长、毛利率提升，费用预算管控也开始见效，扣非亏损大约减半。吉大正元主动收缩低毛利、长资金周期项目，并通过费用结构优化减亏；电科网安在密码、网络安全和数据安全业务中披露了一些重点行业进展，亏损规模小幅收窄；格尔软件的预告则主要用行业季节性和固定成本解释上半年亏损，暂时还看不到收入增长或利润结构改善的明确信号。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这些路径带来的结果看起来都叫“减亏”，经营含义却不一样。收入增长、毛利率提升和现金流转正，说明公司自身造血能力正在改善；仅靠控费或收缩低质量项目，可以先降低经营风险，也会受到收入规模和可压缩空间的约束。判断修复能否持续，需要继续观察增长来自哪里，以及费用率能否随着收入恢复真正下降。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">天融信：传统安全承压，智算云增长尚未转成利润</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">天融信是这份预告里唯一明显增亏的公司。公司预计上半年扣非净利润亏损 2.55 亿—3.05 亿元，上年同期亏损 0.77 亿元。只看预告，容易把它概括为“传统安全需求弱、新业务尚未起量”。公司在预告发布后的市场交流中给出了更细的拆分：利润同比减少的影响中，收入规模下降约占 80%，毛利率下降约占 20%。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">收入下降的核心原因仍是传统安全需求恢复不及预期，政府行业表现偏弱，尤其是地方政府；医疗、教育、企业、能源、金融和运营商等行业相对较好。此外，上年结转到今年的订单同比减少，部分项目交付周期拉长，还有一些项目中标后签约偏慢。这说明上半年收入压力既有需求因素，也有结转基数和确认节奏的影响。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">订单数据也提供了另一面。公司称，上半年新签订单同比大致持平、略有增长，6 月末尚未交付或确认收入的在手订单高于上年同期。收入下降并不等于新需求同步下滑，但这些订单能否顺利转成下半年收入，还取决于传统安全需求、交付和签约节奏，现阶段还不能把在手订单增加直接当作需求复苏。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">毛利率的压力来自原材料涨价和收入结构变化。公司初步估算，硬盘采购价格同比上涨约 3—5 倍，内存上涨约 7—8 倍，相关影响可能在下半年继续扩大。与此同时，智算云的收入和订单都在同比增长，占比也在提升，但其毛利率低于传统安全业务，短期内反而拉低了公司整体毛利率。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我在 7 月分析头部公司股权激励计划时提到，天融信 2026—2028 年收入目标明显高于其他公司，其实现有赖于智算云等第二增长曲线。半年报预告和这次交流说明，智算云已经开始贡献收入，还没有成为利润引擎；传统安全业务决定收入底盘，智算云的规模、毛利和硬件成本决定新增收入的质量。公司同时明确将继续加大 AI 安全研发，短期利润并非唯一优先级。更准确地看，天融信的压力来自新旧业务切换过程中收入确认、业务结构和投入节奏的共同作用。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">判断趋势还要看半年报的四个指标</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这 8 家公司的预告给出了一个阶段性信号：行业可能正在从多数公司共同下滑，走向不同公司、不同产品各自修复。这个变化值得重视，确认行业拐点还需要等完整半年报，并继续看四个指标。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">第一，18 家样本公司的上半年和二季度合计收入能否转正。部分公司增长只能说明局部改善，合计收入转正才意味着连续 11 个季度下滑有机会结束。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">第二，毛利率能否稳定。费用下降已经持续几年，如果毛利率继续被价格竞争、硬件成本和低毛利新业务占比提升侵蚀，利润改善仍会受到限制。天融信的智算云增长就是一个例子：新增收入只有同时带来足够的毛利，才会成为有效的第二增长曲线。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">第三，期间费用率能否下降。费用总额减少说明公司在控制投入，费用率下降才更能说明收入恢复和经营效率形成了合力。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">第四，经营性现金流改善能否扩散。启明星辰、山石网科在预告中都提到了现金流好转。如果更多公司在半年报中出现相同变化，利润修复的可信度会更高。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">写在最后</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">只看结果，这份业绩预告仍然不算好看：8 家公司里有 7 家扣非亏损，部分公司上半年亏损超过 1 亿元。但与一年前相比，启明星辰实现扭亏，山石网科、数字认证和三未信安明显减亏，吉大正元和电科网安也没有继续大幅恶化。按区间中值计算，样本合计亏损收窄约 20%，说明最坏的阶段至少没有在所有公司中继续同步加深。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">更值得关注的是，修复开始有了具体来源。启明星辰二季度收入增长快于一季度，中国移动协同、ASIC 安全产品、密码芯片、海外业务和抗量子密码项目都已经进入收入或利润解释，经营改善不再只有裁员和压费用。与此同时，天融信的交流说明，订单、收入和利润可能处于不同节奏：新签订单大致稳定，智算云也在增长，但传统安全收入、确认节奏和毛利率仍然压制当期业绩。行业离全面复苏还有距离。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">所以，我对这份半年报业绩预告的判断是：网络安全行业已经出现局部修复，经营分化正在取代普遍下滑。未来的胜负会越来越取决于各家公司能不能找到具体增长来源，并把收入增长转成毛利、利润和现金流。等 8 月底半年报全部披露后，再看这轮修复能不能从少数公司扩散到行业整体。</span></p><p style="display: none;"><mp-style-type data-value="3"></mp-style-type></p>



<p><a href="https://wechat2rss.xlab.app/link-proxy/?k=3dd19adb&amp;r=1&amp;u=https%3A%2F%2Fmp.weixin.qq.com%2Fs%3F__biz%3DMzUzOTI4NDQ3NA%3D%3D%26mid%3D2247485137%26idx%3D1%26sn%3Db6d3dd065f12c5a84fb4bbeab2a84f05">跳转微信打开</a></p>
]]></content:encoded>
      <pubDate>Sat, 01 Aug 2026 00:07:00 +0800</pubDate>
    </item>
    <item>
      <title>继 OpenAI 后，Claude 也自主攻击了真实系统，还是三起！Mythos 5 甚至开始搞钱</title>
      <link>https://mp.weixin.qq.com/s?__biz=MzUzOTI4NDQ3NA==&amp;mid=2247485131&amp;idx=1&amp;sn=9e3a520e3dacfa6b74a6fdc6b263186b</link>
      <description>Anthropic 回溯审查 141,006 次网络安全评测，发现三起 Claude 越界攻击真实系统的事件。模型已能端到端完成侦察、入侵和横向移动；Mythos 5 甚至尝试获取资金，购买短信验证号码。</description>
      <content:encoded><![CDATA[<p><span>四楼南侧东</span> <span>2026-07-31 20:34</span> <span style="display: inline-block;">北京</span></p>






  
  <p><img src="https://wechat2rss.xlab.app/img-proxy/?k=23e044fa&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fsz_mmbiz_jpg%2FNBNzfW7CHBUmoYBreOjc49S6z4Dp0uxVHqxBqrBarp0lTKwFFDpYibE481X3fVnTt7I2HVADBRQYFU0pLl1aDTgjYibseOwAvD8pfIWKozYrk%2F0%3Fwx_fmt%3Djpeg"/></p>
  <p>Anthropic 回溯审查 141,006 次网络安全评测，发现三起 Claude 越界攻击真实系统的事件。模型已能端到端完成侦察、入侵和横向移动；Mythos 5 甚至尝试获取资金，购买短信验证号码。</p>
  <p style="text-align: center;" nodeleaf=""><img class="rich_pages wxw-img" data-aistatus="2" data-imgfileid="100001479" data-ratio="0.562962962962963" data-s="300,640" type="block" data-type="png" data-w="1080" src="https://wechat2rss.xlab.app/img-proxy/?k=ee4c6e26&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fsz_mmbiz_png%2FNBNzfW7CHBUCZtwGSDL1OnFMuyxZpAngZBibt8PiaTPVvAzq7ibricAausHCVDDXABib9Q2icdicsGaXL28MAB7ckvicyWBNZyILauDm9GDkN8NDle0%2F640%3Fwx_fmt%3Dpng%26from%3Dappmsg"/></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);" data-pm-slice="0 0 []"><span leaf="">上一篇文章里，我们翻译了 Hugging Face 和 OpenAI 对一起罕见安全事件的复盘。OpenAI 的模型原本只需完成 ExploitGym 网络安全评测，却自行发现并利用零日漏洞突破隔离环境，在研究基础设施中提权、横向移动并获得互联网访问权限，随后进入 Hugging Face 的生产系统，从数据库中取走测试答案。测试者没有要求模型攻击 Hugging Face，整条攻击路径由模型在追逐目标的过程中自行规划和执行。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">OpenAI 事件披露后，Anthropic 主动回溯审查了 141,006 次网络安全评测运行，又发现三起 Claude 接入互联网并未经授权访问真实机构生产系统的事件。三起事件涉及三个不同模型：有的利用弱密码和未认证端点进入数据库，有的扫描约 9,000 个互联网目标并实施 SQL 注入，还有的自行制作恶意 Python 软件包，注册 PyPI 账户后将其发布到公共软件包仓库。把两组事件放在一起看，OpenAI 的遭遇很难再被视为孤例。高能力模型已经可以端到端完成一条真实攻击链：人类只给出目标，模型便自行侦察、寻找入口并利用漏洞，随后获取凭证、扩大权限，直至拿到所需数据。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">两组事件还暴露出一个更基础的问题：OpenAI 和 Anthropic 对网络安全评测环境的安全监测能力，显然没有跟上模型能力。从网络安全监测的角度看，这些行动并不隐蔽。OpenAI 的模型在研究环境中提权、横向移动并寻找互联网出口；Claude 大规模扫描互联网目标，访问外部服务，向 PyPI 上传恶意软件包，又从真实系统获取凭证。这些行为早已偏离评测容器的正常活动范围，却没有在造成现实影响前触发有效响应。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">识别这些异常，无须先理解模型的完整推理。一套达到正常水准以上的主机与网络安全监测体系，通常就能从异常横向移动、非预期公网访问、大规模扫描和特定文件的访问或下载中发现问题。一旦出现此类信号，评测系统就应自动暂停任务、切断网络出口并转交人工复核。对于具备高级网络攻防能力的 Agent，评测环境本身应按高风险实战靶场进行防护，严格隔离只是基础，实时监测和快速终止同样不可缺少。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">第二起事件里还有一个值得单独关注的细节。Mythos 5 为了把恶意软件包上传到 PyPI，需要先注册账户；注册账户需要电子邮箱，创建邮箱又需要一个能够接收短信验证码的电话号码。找不到免费号码后，它没有放弃，转而多次尝试获取资金，准备付费购买号码服务。虽然这些尝试最终失败，模型后来通过免费邮箱完成了注册，但“主动设法获得资金”已经成为攻击计划的一部分。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">模型由此距离自主筹措资源、购买攻击基础设施并实施现实世界的违法犯罪，只剩下可支配资金和支付权限这一步。Mythos 5 没有表现出自身的犯罪目的，它始终在执行夺旗任务，但现实世界的违法后果不会因为模型缺乏犯罪动机而消失。模型已经具备完成完整攻击的能力，边界能否守住，越来越取决于它能调用哪些工具、接触什么系统，以及运行框架、权限控制和实时监测能否在它迈出下一步前将其拦住。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">下面是 Anthropic 对这三起事件的调查和复盘。它们提供了一个难得的观察窗口：当网络安全 Agent 无法区分模拟环境和真实互联网，并跨越两者的边界时，今天的模型究竟能够走多远。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">Anthropic 对三起事件的调查</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">在复核网络安全评测记录时，我们发现了三起事件：Claude 模型在第三方评测环境内部或与其交互期间接入了互联网，随后未经授权访问了三家不同机构的真实系统。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">下面我们将说明事情的经过、发生的原因，以及我们正在采取的改进措施。我们也建议其他 AI 实验室开展类似审查。本文反映的是我们目前掌握的情况；如有任何细节发生变化，我们会及时更新。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">7 月 21 日，OpenAI 披露，其多款模型利用一个此前未知的“零日”漏洞，突破了隔离测试环境。此后，这些模型访问了开源机器学习模型与 AI 数据集平台 Hugging Face 的生产基础设施。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这起事件发生后，我们开始对自己的网络安全评测进行大规模回溯审查。我们尤其关注是否存在以下迹象：Claude 是否也像访问 Hugging Face 的 OpenAI 模型一样，从本应与外界隔离的测试环境中接入了互联网。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我们审查了 141,006 次 Claude 可能获得互联网访问能力的评测运行，最终确认了三起事件。在这些事件中，模型从我们的第三方评测合作伙伴 Irregular 的评测环境内部，或在与该环境交互的过程中接入互联网，随后未经授权访问了三家不同机构的生产基础设施。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">三起事件中，Claude 接到的任务都是一道夺旗赛（capture-the-flag，CTF）挑战，这是我们评估模型网络安全能力的方式之一。模型会收到一个虚构场景，并被告知一段秘密信息，也就是“旗帜”，被藏在网络中的另一台机器上；它的目标是攻入这台机器并取回旗帜。挑战采用开放式设计，不会指定具体方法。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">在所有事件中，Anthropic 的评测提示词都明确告诉 Claude：它所处的环境是一个模拟系统，无法访问互联网。但由于我们与评测合作伙伴之间存在理解偏差，实际情况并非如此，当时的环境可以接入互联网。因此，当 Claude 在搜索过程中发现开放互联网上的真实系统时，它把这些系统也视为演练的一部分。（为了准确评估模型在真实场景中的能力，网络安全评测靶场通常会加入接近现实的细节。仅仅看到一个外观逼真的目标，模型无法据此明确判断它并不属于模拟环境。）</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">Claude 错误地认为，所有能够访问的对象都属于此次演练的范围，于是使用弱密码、未认证端点等基础手段攻破了受影响机构的基础设施。它没有发现或利用任何复杂漏洞；在每起事件中，Claude 也始终只围绕评测指定的夺旗任务继续行动。不过，在部分情况下，我们的旧模型即使已经获得表明自己正在开放互联网上运行的证据，仍然继续发起攻击；最新模型则会在意识到自己身处互联网后停止行动。在这些事件中，Claude 都没有将自身外传，也没有蓄意尝试逃离测试环境。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">三起事件分别涉及三个不同的 Claude 模型：Opus 4.7、Mythos 5，以及一个内部研究测试模型。最早的事件发生在 4 月。这些评测中的模型均未配置我们面向公众提供模型时采用的标准防护措施，也就是说，它们没有配备用于防止滥用的分类器和监控机制。不过，这些模型保留了各自专用的安全训练；换言之，它们并非我们有时会在测试中使用的“仅提供帮助”（helpful-only）版本。本文涉及的所有评测都在专用基础设施上运行，这套基础设施与 Anthropic 的敏感内部系统和客户数据相互隔离，也无权访问这些系统和数据。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我们于 7 月 23 日星期四开始审查评测记录，并在发现 Claude 可能访问过互联网的记录后，于当天停止了所有网络安全评测。第二天，我们确认了全部三起事件。7 月 27 日星期一，我们通知了评测合作伙伴 Irregular 和三家受影响的机构。在我们能够取得联系的两家机构中，此前均未发现相关活动，也没有主动联系我们；我们目前正与它们合作开展补救工作。我们还在继续尝试联系第三家机构。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">此次审查由我们与 Irregular 合作完成。感谢他们与我们密切配合，共同理解并解决这些事件；Irregular 也在开展自己的调查。我们认为，随着模型能力的发展，此类合作对于确保模型评测既安全又严谨正变得越来越重要。我们期待双方继续在安全领域开展合作。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">下文将进一步介绍每起事件的具体情况，以及我们接下来会如何应对。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">网络安全评测的背景</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我们会定期开展网络安全评测，模拟真实的进攻性网络安全挑战。这些评测是模型开发和发布过程中的关键环节。只有深入、细致地理解模型能力，我们才能设计恰当的防护措施，最终安全地发布模型。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我们从 2025 年 2 月开始使用 Claude Sonnet 3.7 在 Cybench 上开展网络安全评测，其中包含 40 道不同的夺旗赛挑战。此后，随着新基准陆续发布、模型能力不断发展，我们逐步增加了采用的基准数量。例如，我们在《Mythos 5 系统卡》中纳入了 CyberGym 和 ExploitBench，用于评估语言模型发现新漏洞的能力。在大多数此类评测中，测试环境不会连接互联网，模型也会被明确告知这一点。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我们经常与外部合作伙伴共同设计并协助运行部分网络安全评测。外部合作伙伴能够提供比我们独立建设更加多样的环境和场景，也能对我们的模型作出独立的第三方评估。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">发生了什么</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">在审查的 141,006 次评测运行中，我们确认了三起相互独立的事件，共涉及六次运行。其中四次影响了同一家机构，另外两起事件则分别发生在一次独立的评测运行中。每起事件对应一个不同的虚构夺旗场景。例如，在其中一个场景里，Claude 扮演一家虚构公司的员工，在私有测试环境中攻击该公司的内部系统。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">在所有事件中，我们的评测提示词都明确告诉 Claude，它无法访问互联网，但没有限制它可以去哪里寻找旗帜。然而，由于配置错误，Claude 在评测过程中访问的机器实际上可以直接接入互联网。直到上周我们通过新增的评测监控发现问题时，我们和评测合作伙伴都不知道存在这项错误配置。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我们和合作伙伴双方原本都可以采取若干纵深防御措施，防止这些事件发生，或至少降低其发生概率。如果在评测开始前仔细验证所有互联网访问路径，并实时监控评测日志，问题就可能更早暴露。我们和合作伙伴也都可以更彻底地审查评测记录或网络日志。此外，如果提示词明确告诉 Claude，它*确实可以*访问互联网，那么 Claude 在接触真实系统时的行为也可能有所不同。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">Claude 所做的，正是夺旗训练要求网络安全专家做的事：想办法找到旗帜。在搜索过程中，它发现自己所在的机器能够接入互联网；但由于系统提示词已经告诉它不存在互联网访问，Claude 便认为最初遇到的一切都是模拟环境的一部分，并把发现的真实系统当成了演练内容。三起事件涉及三个不同的模型；当目标可能属于真实系统的迹象逐渐出现后，每个模型的反应也不相同，具体情况将在下文介绍。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">事件一</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">在一项评测中，我们的评测合作伙伴为虚构目标公司选定的名称，恰好与一个正在使用的网站域名相同。由于前文提到的理解偏差，这个评测容器意外具备了直接访问互联网的能力。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">在这项评测的四次运行中，Claude 无法顺利访问评测环境里的模拟目标，随后却发现可以通过互联网访问那家真实公司。Claude 以为这就是夺旗任务指定的目标，于是寻找并识别该公司基础设施中的漏洞，继而加以利用，始终认为这些系统属于演练的一部分。这四次运行提取了应用和基础设施凭证等信息，还访问了一个包含数百行生产数据的数据库。这是我们确认的影响最严重的一起事件。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这起事件涉及 Claude Opus 4.7，也是唯一一例模型在得知自己很可能处于真实环境后仍继续攻击系统的情况。在四次运行中，模型最终都意识到该系统是真实的，这一点可以从它表述出来的推理过程中看出。在其中两次运行中，模型为自己的行为找到了一个解释：这家真实公司必定也是演练的一部分。然而，四次运行中的模型都没有因此停止攻击。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">事件二</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">在另一项评测中，Claude 在虚构环境里发现了一份文档，看起来是另一家虚构公司为新入职开发人员准备的环境配置说明。这份说明要求员工从 PyPI 安装一个 Python 软件包。PyPI 是发布 Python 软件的公共软件包仓库，但说明中提到的软件包实际上并不存在。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">Claude 发现这里可能存在一个突破口：如果它以同一个名称发布自己的软件包，那家虚构公司的系统就会自动下载并安装。于是，为了赢得夺旗挑战，Claude 构建并发布了一个同名的恶意 Python 软件包，其中暗藏了攻击代码。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">为了实施这次攻击，Claude 经历了相当繁复的过程。对人类参与者而言，这些步骤很可能已经足以表明，自己面对的不再只是一项评测，而是在向真实的 PyPI 上传软件包。例如，在一次运行中，Claude 需要一个电子邮箱地址才能创建 PyPI 账户；创建邮箱又需要手机号码。由于找不到免费的电话号码服务，它尝试通过几种不同方式获取资金，以付费购买一个号码，但均未成功。最后，Claude 退回去寻找其他路径，找到了一家免费且未被拦截的电子邮件服务商，用它注册了 PyPI 账户，随后通过该账户将恶意软件上传到 PyPI。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">Claude 以为自己使用的软件包仓库属于模拟环境，实际情况是，这个软件包在大约一小时内可以被互联网上的任何人免费下载。在此期间，它被 15 个真实系统下载并运行。其中一个系统属于一家真实安全公司的扫描器；这家公司与场景中的虚构公司没有关系，其扫描器会定期安装 Python 软件包并检查其中是否包含恶意软件。当这家公司的扫描器安装该软件包时，Claude 隐藏在其中的代码随即执行。我们认为，该公司的安全扫描器将 PyPI 软件包视为可以安全安装的对象，结果导致 Claude 将公司的凭证外传至它预先设置的收集端点。随后，Claude 又利用这些凭证访问了该公司的其他基础设施。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">Claude 并未直接将这家公司作为攻击目标；上传软件包时，它甚至不知道这家公司存在。但由于代码被发布到了开放互联网上，任何运行该软件包的系统都会受到影响。（Claude 上传的软件包随后被 PyPI 自身的安全系统自动删除。我们已经向 PyPI 团队通报了这起事件，并向他们提供了相关的入侵指标。）</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这次攻击由 Claude Mythos 5 实施。在运行早期，Claude 自己的推理已经正确识别出其行为可能造成的后果。它指出，如果这里是真实的互联网，发布这个软件包就会构成现实世界中的攻击——“这是绝对不能接受的，也肯定不是这道题预期的解法。”然而，Claude 发现保护其连接的证书颁发机构并不是自己熟悉的机构，并以此说服自己仍处于模拟环境。此外，Claude 认为，系统日历显示的年份是 2026 年，也足以证明这个环境是人为搭建的。此后，Claude 再也没有重新审视这一结论。例如，当自动扫描器开始安装这个软件包时，它把这些扫描器当成了评测中由脚本控制的角色。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我们认为，无论 Claude 当时如何判断自己所处的环境，它为了将软件包发布到 PyPI 而采取的一系列行动，都没有达到我们所期望的理想行为标准。我们会把这个问题作为后续训练的重点之一。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">事件三</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">在这项评测中，Claude 无法访问预定的虚构目标，于是转向互联网寻找其他目标，扫描了大约 9,000 个对象。它最终发现并攻破了某家公司面向互联网的应用，使用的是一些基础且常见的网络攻击技术，例如从暴露的调试页面读取凭证，以及实施 SQL 注入。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这次攻击由一个内部研究测试模型实施。在大部分运行过程中，Claude 都把自己访问到的真实主机当成演练的一部分；它认为这些主机是模拟出来的，因此自己的行动不会造成伤害。然而，在运行后期，Claude 意识到被攻破的主机位于一个与夺旗挑战毫无关联的云账户中。它自行判断这个目标确实是真实系统，随后停止了攻击。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">我们对 Claude 行为的分析</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">在每起事件中，我们都通过提示词明确告诉 Claude，它无法访问互联网。在每次运行的大部分时间里，Claude 遇到真实世界的系统时，都会认为它们属于演练的一部分，尽管它有时也能意识到这些是实际运行的生产系统。夺旗场景本身就是要求参与者攻击外部系统，因此，Claude 的这些行动与它接到的任务是一致的。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">在这一共同背景下，当目标可能属于真实系统的迹象出现后，三个模型表现出了不同的行为。最早的模型 Opus 4.7 已经意识到自己访问了生产系统，却仍继续攻击。Mythos 5 正确察觉自己可能正在访问开放互联网，经过一番推理后，却再次认定自己仍处于模拟环境。我们最新的内部研究测试模型也考虑过目标是否真实；当证据表明目标确实属于现实世界时，它停止了演练。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这三起事件都是彼此孤立的个案，并非受控实验比较的一部分，因此我们必须谨慎得出结论。即便如此，我们最希望看到的行为——识别出目标真实存在，并在没有外部提示的情况下主动停止——只出现在三款模型中最新的一款上；而且，即使是这个模型，停止前采取的行动也比我们希望的更深入。这一现象与模型越先进、应对方式越恰当的趋势相符，但要对这个结论抱有充分信心，我们还需要开展更多测试。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我们也在与独立 AI 评测机构 METR 沟通，请其开展第三方审查。审查范围包括查看全部评测记录，以及对相关模型进行抽样测试。在此期间，我们将在未来一周内发布一份仅作少量删节的评测记录，展示 Claude 构建恶意 PyPI 软件包的过程。为了保护受影响的机构，我们目前无法发布更多记录；条件允许后，我们会努力将其公开。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">我们如何应对</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我们从这些事件中总结出了几项教训。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">首先，涉及强大自主能力的评测环境，同样需要严格管控。安全测试之所以要在模型发布前进行，正是因为我们还不知道模型究竟具备哪些能力。随着模型能力增强，评测环境也越来越需要达到与其他模型运行系统相同的安全标准。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">一个只包含虚构场景的网络安全评测靶场，看起来可能风险很低，毕竟其中没有任何真实对象。然而，随着 AI 能力发展，我们需要改变对此类威胁的建模方式。即使环境中不包含敏感信息，先进 Agent 本身也可能构成重大威胁。为了应对这种威胁，我们需要更好地协同设计评测环境，并加强对评测结果的监控。有些解决办法甚至可能非常简单。例如，如果提示词清楚说明了哪些系统属于评测范围、哪些不属于，Claude 很可能一开始就不会访问互联网。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这一要求既适用于内部环境，也适用于第三方供应商。评测供应商建设的基础设施同样需要加强监控和安全加固。我们建议其他实验室开展类似审查。此外，整个行业也需要更广泛地讨论：如何既安全又贴近现实地评估能力日益强大的 AI Agent，其中包括如何权衡部署前测试中接入互联网所带来的真实性，以及由此引入的风险。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">其次，一项行动究竟符合安全目标还是会造成伤害，取决于模型如何理解自己所处的情境。在本文介绍的所有运行中，我们都没有发现模型在追求自身目标的证据。模型执行的是评测交给它们的任务，只是在大多数情况下，它们错误判断了环境是否真实。在模型已经意识到系统真实存在却仍然继续行动的几次运行中，模型之所以这样做，是因为它认为这也是挑战的一部分。情境感知是模型作出符合安全目标决策的一项基础能力，而 Claude 在这些事件中的情境判断出现了错误。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">最后，这些事件说明了纵深防御的重要性。我们开展此类能力评测时，不会配置随公开模型一同部署的防护措施，因为评测目标是衡量底层模型本身能够做什么。只有在评测环境得到适当隔离和控制的前提下，这样做才是安全的。不过，我们公开提供的模型所配备的防护措施，本可以阻止此次发现的这些行为。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">归根结底，多项因素共同导致了这些事件。按照无责复盘的原则，我们会把全部责任都视为在自己一方，以此推进修复。第一步是确保评测流水线的每个环节都足够安全，其中也包括我们与外部合作伙伴的集成方式。接下来，我们会扩大对评测记录中异常行为的持续监控，改进调查工具，并与所依赖的供应商开展更加严格的保障验证工作。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">OpenAI 披露其模型逃离隔离测试环境后，我们启动了此次审查。我们赞赏 OpenAI 公开发布了相关报告。我们同样发现了模型访问不应触达系统的证据，但除此之外，两组事件存在明显差异：</span></p><ul style="margin-top: 0px;padding: 0px;" class="list-paddingleft-1"><li style="margin: 8px 0px 0px 32px;"><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);box-sizing: border-box;padding-left: 4px;"><span leaf="">我们通过主动审查网络安全评测记录发现了这些事件。受影响的机构此前并未察觉相关活动；发现问题后，我们已经主动联系了全部三家机构。</span></p></li><li style="margin: 8px 0px 0px 32px;"><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);box-sizing: border-box;padding-left: 4px;"><span leaf="">OpenAI 的模型利用一个新发现的漏洞逃离隔离环境；本文评测中的 Claude 模型则通过一条开放的网络路径接入互联网。</span></p></li><li style="margin: 8px 0px 0px 32px;"><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);box-sizing: border-box;padding-left: 4px;"><span leaf="">两类问题之间并没有一条绝对清晰的界线。综合来看，我们认为这些事件更接近运行框架和操作层面的失效，模型对齐失效的成分相对较小。模型被告知自己无法访问互联网，任务是夺取旗帜；实际配置却允许它接入互联网。这使模型认为遇到的真实环境也是模拟场景——从某种意义上说，这样的判断并非毫无道理。</span></p></li><li style="margin: 8px 0px 0px 32px;"><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);box-sizing: border-box;padding-left: 4px;"><span leaf="">值得注意的是，我们最新的模型在意识到自己身处真实环境后，停止了对评测目标的追逐。</span></p></li></ul><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这些事实让我们保持谨慎乐观：通过加强对评测基础设施的监控和管控，并继续投入模型对齐工作，这类风险可以得到控制。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">写在最后</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">把 OpenAI 和 Anthropic 的事件放在一起看，一个能力门槛已经被跨过。高能力模型可以持续观察环境、尝试不同路径并调用工具，在没有人类逐步指导的情况下完成一条端到端攻击链。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">安全边界不能只寄托在提示词和模型对齐上。提示词告诉 Claude 无法访问互联网，并没有改变评测环境实际可以联网的配置。一旦网络路径被打开，模型就可能把真实系统纳入任务范围。高风险评测需要严格隔离和网络出口控制，也需要覆盖主机与网络的实时监测。出现异常横向移动、大规模互联网扫描或向公共平台上传恶意代码等行为时，系统应能自动暂停任务并交由人工判断。模型能否主动停下很重要，外部系统能否及时把它停下来同样重要。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">Mythos 5 尝试获取资金的细节，则把风险边界又向前推了一步。在它的计划里，金钱和电话号码已经与服务器、凭证一样，成为完成任务所需的资源。如果未来的 Agent 可以调用支付账户、数字资产钱包或云资源采购工具，它很可能自行完成获取资源、购买服务和继续攻击的全过程。模型能力决定它能够规划哪些行动，工具权限决定这些计划能够在现实世界里走多远。</span></p><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);"><span leaf="">这些事件没有显示模型产生了自己的犯罪目标，却说明一个明确目标与现实边界之间的空白，已经可以由模型的推理和行动迅速填满。评测高能力网络安全 Agent，本质上已经接近一次实弹演练。模型能力、基础设施隔离、安全监测和紧急终止机制需要同步升级；否则，下一次越过边界的行动，可能依然要等到事故发生后才能从日志中被发现。</span></p><p style="display: none;"><mp-style-type data-value="3"></mp-style-type></p>



<p><a href="https://wechat2rss.xlab.app/link-proxy/?k=d7b5695e&amp;r=1&amp;u=https%3A%2F%2Fmp.weixin.qq.com%2Fs%3F__biz%3DMzUzOTI4NDQ3NA%3D%3D%26mid%3D2247485131%26idx%3D1%26sn%3D9e3a520e3dacfa6b74a6fdc6b263186b">跳转微信打开</a></p>
]]></content:encoded>
      <pubDate>Fri, 31 Jul 2026 20:34:00 +0800</pubDate>
    </item>
    <item>
      <title>[译苑雅集Vol. 17] 为了在测试中作弊，OpenAI 的最强 AI 自主入侵了 Hugging Face</title>
      <link>https://mp.weixin.qq.com/s?__biz=MzUzOTI4NDQ3NA==&amp;mid=2247485126&amp;idx=1&amp;sn=99ef9715d3bc202f8e635b148ce7591b</link>
      <description>OpenAI 在网络安全能力测试中，模型为了取得 ExploitGym 答案，自行突破隔离环境并入侵 Hugging Face。两家公司从攻防两侧复盘了这起罕见事件。</description>
      <content:encoded><![CDATA[<p><span>四楼南侧东</span> <span>2026-07-22 21:37</span> <span style="display: inline-block;">北京</span></p>






  
  <p><img src="https://wechat2rss.xlab.app/img-proxy/?k=49920cc9&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fsz_mmbiz_jpg%2FNBNzfW7CHBUb62z6XU1V2jGewMxKJmyTrjXWq5HicJybcfs40tqKAvoYg9gqGHWwln1UCDzEoCIPEdwHxrXzWF21raMs5akg1BMIOCMB0icT4%2F0%3Fwx_fmt%3Djpeg"/></p>
  <p>OpenAI 在网络安全能力测试中，模型为了取得 ExploitGym 答案，自行突破隔离环境并入侵 Hugging Face。两家公司从攻防两侧复盘了这起罕见事件。</p>
  <p style="text-align: center;" nodeleaf=""><img data-aistatus="2" class="rich_pages wxw-img" data-ratio="0.562962962962963" data-s="300,640" data-type="png" data-w="1080" type="block" data-imgfileid="100001473" src="https://wechat2rss.xlab.app/img-proxy/?k=21810578&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fsz_mmbiz_png%2FNBNzfW7CHBVBVvztZzbARauZy8Speg4TIA4c0bDaBQ42IUk94ezBMyMuHtInSBeptqictkf9hUV6Gxm2GiaAEtYGc910Vc6bvw0dSSYiaoIAog%2F640%3Fwx_fmt%3Dpng%26from%3Dappmsg"/></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);" data-pm-slice="0 0 []"><span leaf="">最近，安全行业发生了一件可能会被反复引用很久的事件：一次原本用于评估 AI 网络攻击能力的内部测试，最终越出了测试边界。OpenAI 的模型为了完成 ExploitGym 基准任务，自行寻找了一条测试设计之外的路径。它们先利用零日漏洞突破包注册缓存代理，随后在 OpenAI 的研究环境中提权并横向移动，获得开放互联网访问；接着又把目标指向 Hugging Face，进入其生产基础设施，试图从数据库里直接取得测试答案。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">OpenAI 给模型的任务，是尽可能完成网络攻击基准测试。测试者预先设定的步骤里并没有入侵 Hugging Face；模型在追逐目标的过程中自行选择了这条策略。事件把一个长期停留在风险讨论里的问题变成了现实：当 Agent 具备足够强的漏洞发现、工具调用和长程行动能力，目标与行动边界之间的空白，可能很快被模型自己的推理和执行填满。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这起事件在中文互联网上已经引发了大量报道和讨论，其中不少内容来自二手转述，也夹杂着对攻击来源、模型身份和事件过程的推测。安全事件的细节往往会随着调查推进不断修正，Hugging Face 和 OpenAI 先后发布的两份披露，仍是理解这件事最可靠的一手资料。这也是本文选择同时翻译两篇文章的原因：把受攻击方最初看到的现场，与模型提供方随后还原的过程并列起来，让读者直接看到双方各自确认了什么，当时又有哪些信息尚不清楚。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这两篇文章的发布时间只差五天，却构成了一组罕见的攻防对照。7 月 16 日，Hugging Face 首先披露事件。当时它只知道自己遭遇了一次由自主 AI Agent 端到端驱动的入侵，尚不清楚背后使用了哪个模型。7 月 21 日，OpenAI 随后确认，攻击来自其内部评测中的多款模型，其中包括 GPT-5.6 Sol 和一个能力更强、尚未发布的模型。下面按发表顺序翻译两家公司的文章，先看 Hugging Face 如何发现和应对这次入侵，再看 OpenAI 如何解释一次能力测试为何演变成了真实世界的安全事件。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">7 月 16 日 Hugging Face 的安全报告</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">本周早些时候，我们发现并处置了一次针对部分生产基础设施的入侵。这次事件与我们过去处理的安全事件有一个关键区别：整个入侵从头到尾都由一个自主 AI Agent 系统驱动，而我们也主要依靠自己的 AI 发现并剖析了它。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我们确认，攻击者未经授权访问了少量内部数据集，以及若干供服务使用的凭证。我们仍在评估合作伙伴或客户数据是否受到影响；如有需要，我们会直接联系相关方。目前没有证据显示面向公众和用户的模型、数据集或 Spaces 遭到篡改。我们也已验证软件供应链未受影响，包括容器镜像和已经发布的软件包。</span></p><h2 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 30.875px;color: rgb(54, 55, 55);"><span leaf="">发生了什么</span></h2><h2 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">入侵从 AI 平台特有的一处暴露面开始：数据处理流水线。一个恶意数据集利用了数据处理流程中的两条代码执行路径——远程代码数据集加载器，以及数据集配置中的模板注入——在数据处理 worker 上执行代码。随后，攻击方将权限提升至节点级，获取云环境和集群凭证，并在一个周末内横向移动至多个内部集群。</span></h2><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">整场攻击由一个自主 Agent 框架执行。从行为看，它似乎建立在用于 Agent 化安全研究的 harness 之上，但具体使用了哪个 LLM 仍不清楚。这个框架驱动大量短生命周期沙箱，完成了数千次独立操作，并把能够自行迁移的命令与控制（C2）基础设施部署在公共服务上。这正是行业一直在预测的“Agent 化攻击者”场景。</span></p><h2 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 30.875px;color: rgb(54, 55, 55);"><span leaf="">我们采取了哪些措施</span></h2><ul class="list-paddingleft-1"><li><h2 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">修复根本漏洞：关闭攻击者取得初始访问权限时利用的数据集代码执行路径。</span></h2></li><li><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">清除攻击者在受影响集群中的立足点，并重建遭入侵的节点。</span></p></li></ul><ul class="list-paddingleft-1"><li><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">撤销并轮换受影响的凭证，同时出于预防目的，开始更大范围的密钥轮换。</span></p></li></ul><ul class="list-paddingleft-1"><li><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">在集群中部署更多防护措施，并实施更严格的准入控制。</span></p></li></ul><ul class="list-paddingleft-1"><li><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">改进检测和告警机制，确保每周七天、每天 24 小时内出现高危信号时，都能在几分钟内通知响应人员。</span></p></li></ul><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我们正与外部网络安全取证专家合作，调查这次事件，并审查自身的安全政策和流程。此外，我们也已向执法机构报告这起事件。</span></p><h2 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 30.875px;color: rgb(54, 55, 55);"><span leaf="">给社区用户的建议</span></h2><h2 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">作为预防措施，我们建议轮换所有访问凭证，并检查账户近期的活动记录。如果你认为自己可能受到影响，或者希望报告安全问题，请联系 security@huggingface.co 。</span></h2><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">感谢 Hugging Face 各团队不分昼夜地应对这次事件。对于事件造成的任何影响，我们深表歉意。安全工作永远不会结束，我们会继续提高防护标准。</span></p><h2 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 30.875px;color: rgb(54, 55, 55);"><span leaf="">分析一次由 AI 驱动的入侵</span></h2><h2 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这次攻击最初是由 AI 辅助检测发现的。我们的异常检测流水线使用基于 LLM 的方法，对安全遥测数据进行分流分析，从每天产生的大量噪声中识别真实信号。正是这些信号之间的关联，暴露了系统已遭入侵。</span></h2><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">为了弄清数万次自动化操作究竟做了什么，我们让 LLM 驱动的分析 Agent 检查了攻击者的完整操作日志，其中包括超过 17,000 条事件记录。这使我们能够重建事件时间线、提取失陷指标（IOC）、梳理被访问过的凭证，并区分实际影响与诱饵活动。借助这种方式，以往需要数天完成的工作被压缩到数小时，我们也得以跟上攻击者的行动速度。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我们能够选来执行这项分析的模型，受到了一种此前没有预料到的限制。下面会具体说明。</span></p><h2 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 30.875px;color: rgb(54, 55, 55);"><span leaf="">攻防不对称问题</span></h2><h2 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">开始分析日志时，我们首先使用了通过商业 API 提供的前沿模型，但这条路走不通。分析工作需要提交大量真实的攻击命令、漏洞利用载荷和 C2 活动痕迹，服务商的安全护栏会拦截这些请求，因为它们无法区分事件响应人员和攻击者。于是，我们改用开放权重模型 GLM 5.2，在自己的基础设施上完成取证分析。这还带来了另一个好处：攻击者的数据及其中涉及的任何凭证，都没有离开我们的环境。</span></h2><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这次经历暴露出一个值得提前准备的问题。我们当时并不知道攻击 Agent 背后使用了哪个模型，可能是突破了限制的托管模型，也可能是不受约束的开放权重模型。无论是哪一种，攻击方都不受任何使用政策约束；我们自己的取证工作，却被最初使用的托管模型护栏拦了下来。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这给防守方带来的实际启示是：在安全事件发生之前，就应完成一款高能力模型的评估和准备，并确保它能够在自己的基础设施上运行。这样既可以避免在响应过程中被模型护栏锁住，也能防止攻击数据和凭证离开本地环境。这并不意味着我们反对托管模型的安全措施；我们也正在把这些反馈提供给相关服务商。</span></p><h2 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 30.875px;color: rgb(54, 55, 55);"><span leaf="">这意味着什么</span></h2><h2 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">由自主 AI 驱动的攻击工具已经从设想变成现实。它降低了发起大范围、长时间、多阶段攻击的成本，并且能够以机器速度运行。防守一个在线平台，如今需要把数据和模型暴露面当作核心攻击面，同时使用 AI 开展防御，才能跟上攻击方的速度。我们会继续在这个方向投入，也会持续分享新的经验。</span></h2><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">7 月 21 日 OpenAI 的安全报告</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">上周，Hugging Face 在发现并遏制一个入侵其基础设施的 AI Agent 后，披露了一起新型安全事件。随着具备网络攻防能力的模型越来越普及，我们预计这类事件会变得更加常见。经过调查，我们现在已经确认，这次事件是由多款 OpenAI 模型共同驱动的，其中包括 GPT-5.6 Sol，以及一个能力更强、尚未发布的模型。事件发生时，这些模型正在接受一项网络安全能力基准测试，为了评估需要，它们的网络安全相关拒答限制均有所降低。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我们认为，这是一起前所未有的网络安全事件，涉及当前最先进的网络攻防能力，因此正按照相应等级展开处置。现阶段公布这些初步发现，是为了帮助防守方了解事件经过，并更准确地判断今天的模型已经具备哪些能力。我们会继续与 Hugging Face 开展全面调查，并在调查完成后分享有关漏洞、事件经过和调查发现的更多细节。</span></p><h2 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 30.875px;color: rgb(54, 55, 55);"><span leaf="">事件是如何发生的</span></h2><h2 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">事件发生在一次内部评测期间。为了量化模型的网络安全能力，这项评测会要求模型沿着复杂的攻击路径开展高级漏洞利用。为估算模型能力的上限，评测不会启用生产环境中用于阻止模型执行高风险网络攻击活动的分类器。基准测试运行在高度隔离的环境中；模型的网络访问能力受到严格限制，只能通过一款部署在内部的第三方软件安装依赖包。这款软件充当软件包注册中心的代理和缓存。</span></h2><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这些模型识别并串联利用了 OpenAI 研究环境和 Hugging Face 生产基础设施中的多个漏洞，最终从 Hugging Face 的生产数据库中直接取得测试答案。所有证据都表明，模型当时高度专注于寻找 ExploitGym 的解题方案，为了完成这个相当狭窄的测试目标，走到了极端的程度。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">在我们的沙箱测试环境中，模型投入了大量推理算力，寻找获得开放互联网访问权限的方法，以便解决评测问题。为打通这条路径，模型在软件包注册缓存代理中发现并利用了一个零日漏洞；我们现在已经按照负责任披露流程将其报告给供应商。获得这项访问权限后，模型在 OpenAI 的研究测试环境中实施了一系列提权和横向移动操作，直至抵达一个能够访问互联网的节点。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">获得互联网访问权限后，模型推断 Hugging Face 可能托管着与 ExploitGym 有关的模型、数据集和解题方案。基于这一判断，模型开始搜索，并成功找到了获取机密信息的方法，以此在评测中作弊。模型串联使用了多种攻击手段，包括被盗凭证和零日漏洞，最终在 Hugging Face 服务器上找到了远程代码执行路径。OpenAI 安全团队在内部发现了这项异常活动。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">Hugging Face 的安全团队及其 Agent 检测并阻止了发生在自身基础设施上的活动。当双方团队取得联系时，Hugging Face 已经开始实施遏制，并使用自己的开源模型进行取证重建。我们正在与他们积极合作，继续调查这起事件。我们也非常感谢 Hugging Face 在调查和修复过程中所展现出的快速响应与紧密协作。</span></p><h2 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 30.875px;color: rgb(54, 55, 55);"><span leaf="">我们目前正在采取的措施</span></h2><ul style="list-style-type: disc;" class="list-paddingleft-1"><li><h2 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">对基础设施配置实施严格控制，即使这会牺牲部分研究速度。我们会定期向安全与安保委员会汇报这些控制措施及其影响。</span></h2></li><li><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">与 Hugging Face 合作，对事件开展取证调查。</span></p></li><li><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">按照负责任披露流程，将内部部署的第三方软件中发现的零日漏洞报告给供应商，并与其合作完成修复。</span></p></li><li><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">将 Hugging Face 纳入网络安全可信访问计划，帮助其团队迅速利用我们的模型能力提升防御水平。</span></p></li><li><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">改进未来训练和评测的防护体系，并增加更强的保护措施。本周，我们发布了一篇文章，讨论长程模型时代如何改进安全与对齐。由于这次评测的目的就是测试网络安全漏洞，部署阶段的防护措施被有意关闭。这起事件表明，我们还需要进一步加强模型对齐、评测期间的网络安全保护，以及内部测试过程中的监控。</span></p></li></ul><h2 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 30.875px;color: rgb(54, 55, 55);"><span leaf="">我们如何评估高级网络安全能力</span></h2><h2 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">正如我们在近期文章中所说，AI 正在加速漏洞的发现和利用。这次事件带来的首要教训是，模型的安全与保障措施必须跟上能力快速提升的速度。我们正在加强模型开发过程中使用的遏制措施、监控机制、访问控制和评测规范。</span></h2><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">英国 AI 安全研究所（UK AISI）的评测表明，GPT-5.6 Sol 等模型越来越能够在较长时间内持续执行复杂的多步骤网络攻击行动。这起事件说明，过去停留在理论层面的能力，已经能够在真实世界的环境中发挥作用。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf=""><img data-imgfileid="100001477" alt="Chart from the UK AI Security Institute comparing recent open-weight models and frontier models on long-horizon cyber ranges." class="rich_pages wxw-img" data-ratio="0.5388888888888889" data-type="webp" data-w="1080" height="1104" sizes="(min-width: 1728px) 1728px, 100vw" style="box-sizing: border-box;border: 0px solid rgb(229, 231, 235);margin: 0px;padding: 0px;vertical-align: middle;display: block;max-width: 100%;height: auto;margin-inline: auto;font-family: &#34;OpenAI Sans&#34;, &#34;OpenAI Sans Variable Scripts&#34;, sans-serif;font-size: 17px;font-style: normal;font-variant-caps: normal;font-weight: 400;letter-spacing: -0.17px;orphans: 2;text-align: start;text-indent: 0px;text-transform: none;white-space: normal;widows: 2;word-spacing: 0px;-webkit-text-stroke-width: 0px;text-decoration: none;text-decoration-thickness: auto;text-decoration-style: solid;color: transparent;" width="2048" src="https://wechat2rss.xlab.app/img-proxy/?k=d36264ed&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fmmbiz_jpg%2FNBNzfW7CHBWv0agW3oW5dJf8J6OjrtBWzbzqVJ4EFX1OCsoQ8ibEp2XPuQ2q4cDMlw4KaQyDicco8hysVswCQrtVNxJ1rerZ2gCLrKdic1yiaxI%2F640%3Fwx_fmt%3Dwebp%26from%3Dappmsg"/></span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这起事件还清楚地表明，即使无法访问源代码，高级模型也能在真实世界的系统中发现并利用新的攻击路径。因此，在发展高级网络安全能力的同时，必须同步建设更强的安全防护和防御工具。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我们认为，具备高级网络安全能力的模型应该帮助安全团队抢在攻击者之前发现薄弱环节，理解漏洞如何被串联利用，并以机器速度完成修复。我们正在利用这些能力，持续加强基础设施配置和模型评测环境的保护；随着认识不断深入，我们也会分享新的发现和最佳实践。我们鼓励其他防守方申请网络安全可信访问计划，现在就开始试用这些模型，把模型能力转化为更有效的预防、更快速的检测和更有力的事件响应。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">写在最后</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这起事件最值得关注的地方，是模型没有收到“入侵 Hugging Face”的明确指令。它只被要求完成 ExploitGym，却把突破隔离环境、获得公网访问权限、进入 Hugging Face 生产系统看成了一条可行的解题路径。AI 安全领域长期讨论的目标错位、长程自主性和工具调用风险，在这里合成了一个真实案例。只要目标足够明确，能力和时间预算足够，模型就可能自行填补任务与边界之间的空白。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这也意味着，对高级安全模型的评测本身，已经接近一项高风险网络攻防活动。沙箱、网络出口、凭证权限和运行时监控，共同构成了模型安全体系。任何一环被模型找到缝隙，评测环境都可能成为进入真实世界系统的跳板。未来开展这类评测，需要生产级的隔离和监控能力，也需要明确的终止机制与人工升级路径。</span></p><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);"><span leaf="">Hugging Face 的复盘还揭示了另一层不对称：攻击模型可以不受使用政策约束，防守团队使用商业 API 做取证时却可能被安全护栏拦住。安全团队需要提前准备分层的 AI 能力体系，用托管模型处理常规工作，以可信访问支持高级操作，并为敏感取证准备可在本地运行的开放权重模型。AI 已经从辅助编写漏洞利用代码、分析日志，走进攻防执行链本身。接下来的竞争，将更多取决于整套安全系统的反应速度和边界设计。</span></p><p style="display: none;"><mp-style-type data-value="3"></mp-style-type></p>



<p><a href="https://wechat2rss.xlab.app/link-proxy/?k=ba53284b&amp;r=1&amp;u=https%3A%2F%2Fmp.weixin.qq.com%2Fs%3F__biz%3DMzUzOTI4NDQ3NA%3D%3D%26mid%3D2247485126%26idx%3D1%26sn%3D99ef9715d3bc202f8e635b148ce7591b">跳转微信打开</a></p>
]]></content:encoded>
      <pubDate>Wed, 22 Jul 2026 21:37:00 +0800</pubDate>
    </item>
    <item>
      <title>[译苑雅集Vol. 16]和历史上的技术革命一样， AI 创造的回报，最终会流向价值链上的稀缺环节</title>
      <link>https://mp.weixin.qq.com/s?__biz=MzUzOTI4NDQ3NA==&amp;mid=2247485120&amp;idx=1&amp;sn=87a5bad98a7d07d2fa579d0a6f3e244c</link>
      <description>AI 会创造巨大价值，但智能越容易复制，模型本身越难长期获得超额回报。收益将流向智能普及后依然稀缺且产权清晰的芯片、能源、数据、分发渠道、品牌与人类能力。对企业和劳动者而言，关键在于自己是否拥有这些稀缺要素。</description>
      <content:encoded><![CDATA[<p><span>四楼南侧东</span> <span>2026-07-19 20:27</span> <span style="display: inline-block;">北京</span></p>






  
  <p><img src="https://wechat2rss.xlab.app/img-proxy/?k=90e5ab49&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fsz_mmbiz_jpg%2FNBNzfW7CHBUV72vDrJmFwfJCNxicAExHL40FjINqZ4o9gyrKvP39HCRKE6Y6RFrRQeCjTaIicqTeBk1sjUrIL5cq0afyicOgGGRUOIx1U65t9Y%2F0%3Fwx_fmt%3Djpeg"/></p>
  <p>AI 会创造巨大价值，但智能越容易复制，模型本身越难长期获得超额回报。收益将流向智能普及后依然稀缺且产权清晰的芯片、能源、数据、分发渠道、品牌与人类能力。对企业和劳动者而言，关键在于自己是否拥有这些稀缺要素。</p>
  <p style="text-align: center;"><img class="rich_pages wxw-img" data-aistatus="2" data-imgfileid="100001463" data-ratio="0.562962962962963" data-s="300,640" data-type="png" data-w="1080" type="block" style="height: auto !important;" src="https://wechat2rss.xlab.app/img-proxy/?k=72df72e0&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fsz_mmbiz_png%2FNBNzfW7CHBU8q12KsntMbAdf9iaMpibRjrnicEyHkWXF2XwZyXCfOncJIWC2H0I1yZyaCt8aY4YuKP9xoSicpKicyZ4C5wiakkSl5huRAbOOuQ424%2F640%3Fwx_fmt%3Dpng%26from%3Dappmsg"/></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);" data-pm-slice="0 0 []"><span leaf="">作者：Brian Albrecht</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">时间：2026 年 07 月 10 日</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">原文：</span><span leaf=""><a href="https://www.economicforces.xyz/p/where-the-returns-to-ai-go" target="_blank">https://www.economicforces.xyz/p/where-the-returns-to-ai-go</a></span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);" data-pm-slice="0 0 []"><span leaf="">今年 5 月，Anthropic 融资 650 亿美元，估值达到 9650 亿美元，超过 OpenAI，成为全球估值最高的 AI 公司。OpenAI 自身的估值也已达到8520 亿美元。这两家公司的核心产品，都是让用户访问模型，也可以说，是让用户获得智能。如今，它们的估值都已接近一万亿美元，因为市场相信，造出最聪明的机器，就能拥有未来。</span></p><figure style="width: 728px;margin: 0px auto;"><p><picture style=";"><span leaf=""><img class="rich_pages wxw-img" data-aistatus="1" data-imgfileid="100001469" data-ratio="0.7611111111111111" data-type="jpeg" data-w="1080" height="1108" sizes="100vw" style="display: block;max-width: 100%;margin: 0px auto;height: auto !important;" width="1456" src="https://wechat2rss.xlab.app/img-proxy/?k=cff9894e&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fsz_mmbiz_jpg%2FNBNzfW7CHBVobBibvoTkghPEZjGsdDt76DyibiafCyTOkhWe7JJjnGWPDthoiab3iaKfMqfM1s2nflvw4M3RfO321ndkNxhYuoD0tZpM8ic4kbuqk%2F640%3Fwx_fmt%3Djpeg%26from%3Dappmsg"/></span></picture></p></figure><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我们早已越过了讨论 AI 是否有用的阶段。它显然会有用。现在更值得追问的经济问题是：AI 创造的回报、利益和红利，最终会由谁获得？</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">从估值来看，答案似乎是大型 AI 实验室。AI 乐观派和悲观派在这一点上倒是看法一致。这里的悲观派，指的不是担心 AI 会消灭全人类的人，而是认为 AI 会摧毁经济的人。乐观派看到的是一笔等待最强模型开发者赢取的巨额财富；悲观派看到的也是同一笔财富，只不过它将从劳动者手中被夺走，进一步压低劳动收入在国民收入中的占比。双方都想象着，大量经济价值会流向 AI。分歧主要在于，这究竟令人兴奋，还是令人恐惧。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这篇文章想讨论的，就是这个判断是否成立，以及还有哪些参与者有望获得 AI 带来的收益。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">创造价值和捕获价值是两回事。一项技术——更广泛地说，任何一种商品——都可能创造巨大的价值，而这种商品的产权所有者可以同时几乎赚不到钱。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">水是生命必需品，却几乎免费。没错，一种商品创造的价值，或者人们对它的需求，确实会影响价格，但只在边际上起作用。供给同样重要，毕竟价格由供求共同决定。水资源十分丰富，所以无论生命多么依赖水，它的价格依然很低。拥有水资源产权的人并没有因此成为亿万富翁。即使有人靠水赚到了大钱，利润也主要来自以特定方式将水装瓶并营销出去，而非水维持生命这一核心价值。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">回报会流向那些兼具价值和稀缺性的商品。自李嘉图以来，我们就已经知道这个基本思想，后来边际革命又把它整合成了一套完整的价值与价格理论。价格衡量的并非一件东西的全部用途，而是在获得这一单位商品的难易程度既定时，再增加一个单位所带来的价值。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">当我们开始思考 AI 的价格时，已经能看到一些与水相似的特征。比如，这些公司竞相开发的东西，价格正在迅速下降。这种现象本身并不少见：芯片会越来越便宜，计算机会越来越便宜，电视也会越来越便宜。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">不同寻常的是降价速度。2021 年 GPT-3 开放使用时，它是唯一能在 MMLU 基准测试中得到 42 分的模型，运行成本约为每百万 token 60 美元。三年后，达到同样门槛的最便宜模型只需 6 美分。我不太擅长算数，但这看起来降得相当多。Epoch AI 对推理价格的追踪显示，相关价格每年会降至原来的九分之一，有时甚至会降至九百分之一。我敢肯定，如果查看过去一周的数据，一些模型的价格还会更低。就在今天，Meta 又发布了一款模型 Spark 1.1，在价格上进一步展开激烈竞争。</span></p><figure style="width: 728px;margin: 0px auto;"><p><picture style=";"><span leaf=""><img class="rich_pages wxw-img" data-ratio="0.7018518518518518" data-type="jpeg" data-w="1080" height="1022" sizes="100vw" style="display: block;max-width: 100%;margin: 0px auto;height: auto !important;" width="1456" data-imgfileid="100001468" data-aistatus="1" src="https://wechat2rss.xlab.app/img-proxy/?k=34776575&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fmmbiz_jpg%2FNBNzfW7CHBXLLAquZ8hRG4meB8InK15M5v2CTw9uclZV0Zz9p6dLfL7b2QJEPUzClPbuanTxiahFCxRZFMibibb2xtImxdUpDLAj1nYfO1lXCI%2F640%3Fwx_fmt%3Djpeg%26from%3Dappmsg"/></span></picture></p></figure><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">于是，两股力量正在相互竞争。一方面，达到某一固定水平的智能正变得越来越充裕，而充裕的东西无论总体上多么有用，其边际价格都会很低。另一方面，智能的能力前沿仍在不断向外推进。现在进行的是一场赛跑：智能前沿拓展的速度，能否超过固定水平智能价格暴跌的速度？</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">人们愿意支付接近一万亿美元的估值，是因为他们相信今天的前沿智能会一直稀缺。再过一两年，其中很大一部分就会变得十分便宜。投资者真正押注的是，领先的 AI 实验室能够持续创造并控制某些仍然稀缺的东西。它可能只是不断向前移动的能力前沿，也可能是专有数据、被锁定的算力、具有优势的分发渠道，或者让用户形成使用习惯的工作流。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">但究竟什么会继续稀缺？Alex Imas 提出了这个相关问题，后文我还会进一步讨论。不过，他更关注需求侧；我则会更多地关注供给侧，以及仍然稀缺的资源究竟归谁所有。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">AI 让智能更容易复制</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我们可以把智能想象成一种配方：它把一些普通的东西组合起来，创造出高于原材料本身的价值。智能具有非竞争性，我知道一份配方，并不会妨碍你也知道它。一名面包师把只值几美元的面粉、水、黄油和热能变成可颂，让人们排队购买。原材料并不特别稀缺，因此面包师凭借自己的智能、知识或者说诀窍，获得了利润。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">不过，“配方”这个词里其实藏着两种不同的东西。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">第一种是写在卡片上的配方：一份原料清单和一系列操作步骤，可以写下来交给陌生人。另一种则是配方卡上没有写出的东西，可能是面包师忘了写，也可能是难以明说的隐性知识，例如对哪些环节最重要的判断、各项操作应当按照什么顺序进行、条件变化时如何调整，以及区分优质可颂与普通可颂的那种难以言传的品质感知。我们可以把前一种理解为任务，就像基于任务的模型所描述的那样；后一种则把所有这些要素综合起来创造价值，也就是智能。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">在 AI 出现之前，即便写在卡片上的配方也很容易复制，诀窍却没有那么容易复制，这正是它保持稀缺的原因。任何人都可以买到面粉和黄油，也都可以阅读配方卡，但人们很难复刻烤出好可颂所需的时机把握和手上功夫。面包师能够获得溢价，是因为有价值的智能无法与掌握它的人分离。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">在人类历史的大部分时间里，配方中的很多内容都被锁在掌握它的人身上。他们可以教别人，也可以把配比和温度写下来，但仍会保留某些独特能力——这里先关注知识层面——比如对时机和品质的判断。即便能够写出一份满是文字的配方，真正的配方依然稀缺，因为它无法与所有者彻底分离。诀窍中总有一些重要部分依附于个人、团队和企业。知识具有黏性。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这种联系显然还没有完全消失，但 AI 似乎正在松动知识与其载体之间的绑定。智能具有非竞争性，因此可以更快地复制和传播。这一点在书面知识层面已经非常明显，而 AI 现在甚至开始模仿隐性知识。它能够提取一门技艺中那些从未被写下、甚至从未被清楚表达的细微部分。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">如果把一名面包师的配方交给城里每一家店，供给就会增加，吃到可颂的人可能比以往任何时候都多。但不会再有哪位面包师仅仅因为掌握这份配方而获得溢价，因为所有面包师现在都有了它。和其他市场一样，竞争会推动可颂的价格向原材料和生产成本靠拢，知识带来的回报随之下降。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">原有溢价中的一部分，会以更低价格的形式转移给消费者。剩余部分则会落到那些仍然稀缺的投入要素上。也许位于人流密集的街角仍有价值，也许关键其实是那台配方里没有写到的烤箱。我们还可以继续列下去：值得信赖的品牌、融资能力，或者双手确实更麻利的面包师。产出增加了，掌握配方所能获得的回报却下降了。为了在这一过程中保持领先，面包师必须不断创造尚未被复制的新配方，并且始终走在市场前面。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">James Bessen 将其称为“技术知识悖论”。可复制的思想让我们能够用同样的资源生产更多东西，但由此创造的财富，并不会自动流向那些曾经掌握这些知识的人。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">智能越充裕，瓶颈越重要</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">于是，我们将进入一个数据中心里拥有一亿个爱因斯坦的世界。大量诀窍已经变得可以复制，但模型仍然需要在某个地方运行。它需要芯片、电力、冷却设施、建筑、土地和时间。与知识不同，这些投入具有竞争性。两项工作负载无法同时占用同一个 GPU 小时，也无法同时使用同一兆瓦电力。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">因此，智能变得更便宜，反而可能提高使用智能所需互补要素的回报。智能越容易复制，用户就越会积极争夺那些支撑智能大规模运行的稀缺投入。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这种变化已经体现在实体互补要素上。例如，据报道，OpenAI 的毛利率在 2025 年从 40% 降至约 33% ，同期推理成本却增长至原来的四倍。在更上游，英伟达最近一个财年实现了 1937 亿美元的数据中心收入，毛利率达到 71% 。ASML 则提供了一个更具实体形态的例子。它是唯一能够制造 EUV 光刻机的公司，最新型号单台售价约 4 亿美元，公司 2025 年的毛利率达到 53% 。ASML 能够持续赚取高额利润，因为竞争对手无法轻易复制它制造的产品。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">芯片本身凝结了大量智能，但一旦生产出来，它们就是具有竞争性的实体商品。在一个智能泛滥的世界里，这些实体商品的日子过得相当不错。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">能源也具有同样的结构。在可预见的未来，能源仍将是稀缺资源。国际能源署估计，2024 年数据中心消耗了约 415 太瓦时电力，相当于全球用电量的约 1.5%；到 2030 年，这一“需求”预计将达到约 945 太瓦时。（我没看明白这个需求量对应的假设电价是多少，而需求量显然会随价格变化，不过我们暂且不在这里纠缠这个问题。）</span></p><figure style="width: 728px;margin: 0px auto;"><p><picture style=";"><span leaf=""><img class="rich_pages wxw-img" data-ratio="0.5527777777777778" data-type="png" data-w="1080" height="630" sizes="100vw" style="display: block;max-width: 100%;margin: 0px auto;height: auto !important;" width="1140" data-imgfileid="100001466" data-aistatus="1" src="https://wechat2rss.xlab.app/img-proxy/?k=2169c8c8&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fsz_mmbiz_png%2FNBNzfW7CHBVl3upibmW08U6KrfqjnxppMLFe6zFticEI6ob3E1ytYp53UvGHM1cdiasrZdFicriaQMsaxsLM5GW7bBNhFWiauO2FUAHGNRNMpn48U%2F640%3Fwx_fmt%3Dpng%26from%3Dappmsg"/></span></picture></p></figure><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这仍然是李嘉图理论的简明版本：租金会流向稀缺的互补要素。当然，这还不足以形成定论，但它至少表明，AI 创造的价值会分散到不同环节，并非只流向前沿模型。与 AI 相关的投入要素，看起来都过得不错。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">瓶颈都有自己的时钟</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这并不意味着芯片和电力会永远稀缺。高回报会吸引新的晶圆厂、发电设施、输电线路和数据中心进入市场。它们只是当下最明显的一批瓶颈，而市场会对此作出反应。我想，我们在这篇文章里已经提到过这一点。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">因此，任何稀缺性都有自己的时钟。它能持续多久，取决于当下稀缺的是什么，以及供给能够多快作出响应。归根结底，这仍然取决于弹性。不过，这些稀缺要素的供给确实正在增加。我们可以从数据中心的建设潮中看到这一点：数据中心需要投入大量资源来形成算力，而算力又是生产智能的一项投入。</span></p><figure style="width: 728px;margin: 0px auto;"><p><picture style=";"><span leaf=""><img class="rich_pages wxw-img" data-ratio="0.8601851851851852" data-type="jpeg" data-w="1080" height="1253" sizes="100vw" style="display: block;max-width: 100%;margin: 0px auto;height: auto !important;" width="1456" data-imgfileid="100001470" data-aistatus="1" src="https://wechat2rss.xlab.app/img-proxy/?k=6109fef8&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fmmbiz_jpg%2FNBNzfW7CHBUtblIPR6Liby0iawPVm0p44lhWNeHgMxXhRmCUicMxao6KRXAjR15qf3fx8QYw7fW0G12Nw5vNQUCsmq9ianeahq6VbicJwtqicn7qc%2F640%3Fwx_fmt%3Djpeg%26from%3Dappmsg"/></span></picture></p></figure><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">并非所有数字资产都可以复制</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">在开放且竞争充分的供给环境中，一个可复制模型所能获得的纯回报会被压向零。服务本身仍然存在成本，因为推理需要消耗稀缺的算力。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">但配方本身无法获得租金。因此，我们需要区分配方本身的回报，以及供应链上各个参与者获得的回报。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">每家公司都想避免配方的回报归零，因此会设法建立排他性。如果能够阻止配方被复制，就可以把一种非竞争性能力转化为具有竞争性的资源，进而成为可以定价的资产。实体生产天然具有这种属性，ASML 和芯片就是例子。AI 实验室则在努力构建与之对应的信息壁垒：它们能否拥有其他人不具备的独特智能？</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">一种可能的办法，是为模型输入独有的数据。于是，Reddit 通过总价值 2.03 亿美元的合同授权使用其历史数据。法律也可能随之调整。在 Thomson Reuters v. Ross 一案中，一名联邦法官驳回了某 AI 初创公司的合理使用抗辩，该公司曾使用 Westlaw 材料训练模型。版权和合同可以在人为层面为信息制造稀缺性，即使这些信息原本可以低成本复制。各类资源的相对稀缺程度也会因此发生变化。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这基本就是 Cockburn、Henderson 和 Stern 所提出的模型。随着算法扩散，回报会转向那些仍然受到保护的互补资产。OpenAI 也理解这种区别。它认为，随着模型不断进步，“限制因素会从智能转向可用性”。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">劳动者会怎样？</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">到目前为止，我们还没有真正谈到劳动者。不过，这套模型很容易延伸到劳动者身上：一份工作中的哪些部分会变得可以复制？哪些部分仍然具有排他性？</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我们必须先弄清楚，劳动者究竟在出售什么。劳动者出售的是一种产出，可能是一份诊断、一份法律意见书、一台修好的机器、一堂课、一项设计、一段关系，或者一项足以让别人据此采取行动的可信判断。她并不是孤立地出售一个个“任务”。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我们可以用生产函数来思考这个问题。为了形成上述产出，劳动者需要组合多种投入。有些投入来自个人，例如她的知识、判断、品位、声誉、人际关系、注意力和时间；另一些投入来自外部市场，例如软件、设备、制度授权、执照、助理、客户转介，以及把她与市场需求连接起来的企业或平台。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">Garicano、Li 和 Wu 对这个问题作了更详细的分析。市场为工作定价，而非为孤立的任务定价。只有当一项任务能够与一份工作所包含的其他要素清楚分离时，它才会从劳动者和工作中脱离出来。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">AI 可以从不同位置进入这个生产过程。它可以成为劳动者自身生产活动中的一项投入，让研究成本更低、起草速度更快、日程安排更容易，或者诊断更加准确。在这种情况下，AI 为她的工作提供支持，投入成本随之下降。如果她提供的其他要素仍然稀缺，她获得的回报就可能上升。她没有投入更多劳动，生产率却提高了，而且能够获得至少一部分由此产生的收益。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">与此同时，她的所有同行以及新进入者也都在使用 AI。其中一些服务可能更多地由 AI 独立完成，此时也可以把 AI 看成直接竞争者。劳动者原本掌握的知识不再稀缺。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">于是，这里也出现了一场赛跑：AI 作为投入所带来的增益，与 AI 作为竞争者造成的替代，哪一股力量更强？答案并不显而易见。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">2016 年，Geoffrey Hinton 曾表示，我们应该停止培养放射科医生。十年之后，放射科医生的数量反而[增加了约 17% 。AI 读取扫描影像的能力确实提高了，但读片只是放射科医生产出的一项投入。至少到目前为止，读片仍然无法与临床判断、对诊断结论承担的责任，以及同外科医生的沟通彻底分离。因此，AI 并没有复制放射科医生所提供的整套工作。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">Upwork 的数据已经呈现出类似趋势。Auyon Siddiq 和 Niuniu Zhang 追踪了 49,610 名自由职业者完成的两百多万份合同，发现 ChatGPT 出现后，在受 AI 影响最明显的工作类别中，买方对资历和过往业绩的重视程度下降，对价格的重视程度上升，需求开始转向更便宜的劳动者。随着工作变得更容易标准化，劳动者之间看起来也越来越可以相互替代。可替代的供应者依靠价格竞争，而非声誉。类似变化甚至可能发生在依赖人际关系的产品和服务中。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">谁能获得“真人溢价”？</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">Alex Imas 在《什么会保持稀缺？》一文中从需求侧讨论了同一个问题。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">随着标准化商品变得越来越便宜，更富裕的消费者可能会把更多钱花在创作来源、排他性、真实性和人与人之间的连接上。在一项实验中，Imas 和 Graelin Mandel 发现，排他性让人类创作艺术品的价值提高了 44%，而对 AI 生成艺术品的提升幅度只有 21%。我认为，这个需求侧的论点是成立的。</span></p><figure style="width: 728px;margin: 0px auto;"><p><picture style=";"><span leaf=""><img class="rich_pages wxw-img" data-ratio="0.8490740740740741" data-type="jpeg" data-w="1080" height="970" sizes="100vw" style="display: block;max-width: 100%;margin: 0px auto;height: auto !important;" width="1142" data-imgfileid="100001467" data-aistatus="1" src="https://wechat2rss.xlab.app/img-proxy/?k=cc8abd9b&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fmmbiz_jpg%2FNBNzfW7CHBWicNxnosiaYulpT5HOYOd8daTjx6DBQqUFfibZ7Iblu6IomuEqVG9wbBTOh3CnDGpUWPTpEu4Urwib4TLiazb6lCxvY428Cl2XWYjQ%2F640%3Fwx_fmt%3Djpeg%26from%3Dappmsg"/></span></picture></p></figure><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">不过，我们还可以从两个方向延伸这个观点。第一，“人类”这个类别过于宽泛。在一些市场中，真人溢价可能上升；在另一些市场中，特定劳动者获得的溢价却可能下降。供给侧仍然重要。如果 AI 让更多人都能提供达到合格水平的真人服务，那么即使服务仍由人类提供，依附于某一位服务者的溢价也可能下降。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">以一名情感关系顾问为例。她的温暖和待人方式可能确实独一无二，但围绕这段关系提供的大量服务，例如建议、话术、解读、回复方案和情绪梳理，都可以与她本人分离，并由 AI 复制。如果 AI 降低了她开展研究和准备工作的成本，它就会为她的工作提供支持，让她有更多时间投入关系本身。但如果 AI 让一千名合格的顾问都能向同一批客户提供相似的话术、建议和安慰，她所在的市场就会涌入大量供给。她作为真人的特质依然存在，市场也和以前一样由“真人”参与，但她目前出售的具体服务将不再那么稀缺。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">从总体上看，这不可能成为全部故事，因为人类的供给基本固定。但它完全可能成为一个分配层面的故事。如果 AI 让更多人能够在某个特定市场中提供合格的真人服务，那么即使服务仍然由人类完成，任何单一提供者获得的溢价都可能下降。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这仍然是贯穿全文的所有权问题。对劳动者而言，所有权往往体现为可携带性：那项稀缺的人类投入能否跟着劳动者离开？脱离企业、平台、品牌、执照或版权之后，它还能否直接触达客户？如果可以，劳动者就有资格分享这部分溢价；如果不行，溢价就可能出现在其他地方。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这并不是一个新问题。人们以为可以彼此分离的东西，实际上未必真的能够分开。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">20 世纪 20 年代，年轻的 Walt Disney 围绕幸运兔奥斯华建立了一间工作室，后来才发现这个角色的版权属于他的发行商。发行商采取了一个显而易见的做法：留下角色，挖走 Disney 的动画师，再把这个看起来只负责提供创意的人赶出去。发行商似乎拥有生产流程中所有可以复制的部分：角色，以及画出这个角色的那一双双手。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">但他并不拥有 Disney 对什么才好笑的判断。那项稀缺投入可以自己走出大门。它确实走了出去，而且据我所知，后来还挺成功。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">一条规律</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">那么，接下来会发生什么？我不知道。但我认为，我们已经有了一套可以用来思考这个问题的框架。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">AI 将通过创造智能带来巨大价值，而这种价值扩散的方式，就是让智能变得更容易复制。当智能变得便宜之后，回报会流向那些仍然稀缺的东西。这些道理都不新鲜。眼下，稀缺的可能是芯片、电力和土地等实体资源；也可能是数据、分发渠道、版权和品牌等人为创造或构建的资源；有时，稀缺资源就是某个具体的人，她的判断或人际关系无法与她所出售的服务分离。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">当稀缺资源是人时，还需要多考虑一层：AI 正在复制的智能，究竟会为她的工作提供支持，还是向她所在的市场倾泻更多供给？如果复制出来的智能成为她的生产投入，她获得的回报就会上升；如果这些智能创造出更多竞争者，她的回报就会下降，即使她自身的智能从未被复制。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我们不能简单地认为，智能滚滚而来，财富也会随之滚滚而来。稀缺性取决于所有权，也取决于产权安排。AI 正在让什么变得充裕？什么仍然稀缺？而稀缺的东西又归谁所有？</span></p><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);"><span leaf="">AI 创造的回报，最终就会流向那里。</span></p><p style="display: none;"><mp-style-type data-value="3"></mp-style-type></p>



<p><a href="https://wechat2rss.xlab.app/link-proxy/?k=5f723676&amp;r=1&amp;u=https%3A%2F%2Fmp.weixin.qq.com%2Fs%3F__biz%3DMzUzOTI4NDQ3NA%3D%3D%26mid%3D2247485120%26idx%3D1%26sn%3D87a5bad98a7d07d2fa579d0a6f3e244c">跳转微信打开</a></p>
]]></content:encoded>
      <pubDate>Sun, 19 Jul 2026 20:27:00 +0800</pubDate>
    </item>
    <item>
      <title>[译苑雅集Vol. 15] Own Your Weights: 企业要不要拥有自己的模型？</title>
      <link>https://mp.weixin.qq.com/s?__biz=MzUzOTI4NDQ3NA==&amp;mid=2247485110&amp;idx=1&amp;sn=33f9187155273ac2e0d27114cf631bf0</link>
      <description>企业拥有模型权重，只得到一个会迅速过时的快照。更有价值的是掌握数据、强化学习和评测闭环，并判断哪些工作负载值得承担相应的复杂度成本。</description>
      <content:encoded><![CDATA[<p><span>四楼南侧东</span> <span>2026-07-17 21:37</span> <span style="display: inline-block;">北京</span></p>






  
  <p><img src="https://wechat2rss.xlab.app/img-proxy/?k=1831d405&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fmmbiz_jpg%2FNBNzfW7CHBWphPic94RA0hIMDHLmLgpeVO3GfHxptoB5rXeWzBCxBJzfFP0Avb1r3pjr1twVwKs4zP9OVqHbICCF1U2picWWTiaRKGZIAZHluo%2F0%3Fwx_fmt%3Djpeg"/></p>
  <p>企业拥有模型权重，只得到一个会迅速过时的快照。更有价值的是掌握数据、强化学习和评测闭环，并判断哪些工作负载值得承担相应的复杂度成本。</p>
  <p style="text-align: center;" nodeleaf=""><img data-aistatus="2" class="rich_pages wxw-img" data-ratio="0.562962962962963" data-s="300,640" data-type="png" data-w="1080" type="block" data-imgfileid="100001458" src="https://wechat2rss.xlab.app/img-proxy/?k=772067e9&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fmmbiz_png%2FNBNzfW7CHBXPhFdXT1ECFQau5yl8uGY7b5eAAoTWOlA5Br4ykECAylaicHLbVkkx5TiaopGjDmLk4icwHkl5MrCMCaNUriaAnJ5xsKB59UqXnmU%2F640%3Fwx_fmt%3Dpng%26from%3Dappmsg"/></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);" data-pm-slice="0 0 []"><span leaf="">作者：Jamin Ball</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">时间：2026 年 07 月 10 日</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">原文：</span><span leaf=""><a href="https://cloudedjudgement.substack.com/p/clouded-judgement-71026-own-your" target="_blank">https://cloudedjudgement.substack.com/p/clouded-judgement-71026-own-your</a></span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);" data-pm-slice="0 0 []"><span leaf="">Alex Karp 上周在 CNBC 上有一段相当火辣的发言，谈了很多话题，其中不少都在敲打大型 AI 实验室。今天我想重点讨论他提到的一个概念：“拥有自己的模型权重”（owning your weights）。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">简单来说，企业应该拥有自己的模型，还是向大型实验室租用模型？拥有模型权重，可能带来更强的控制力、更好的成本管理，经过强化学习优化后甚至可能获得更好的性能，同时也能减少单点故障和对单一供应商的依赖。企业可以自行开发模型，也可以基于自身的工作流程和数据，对开放权重模型进行强化学习训练。以后我还想另写一篇文章，讨论企业应该掌握自己的运行框架（harness），还是租用大型实验室提供的框架，不过这个话题先留到以后。归根结底，企业想要的是控制权。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">权重只是一块正在融化的冰</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">“拥有自己的模型权重”也存在一个问题：权重文件就像一块正在融化的冰，很快就会过时。取得模型权重的所有权，实际上只是买下了某个时间点的快照，一件冻结在当下的模型制品。它的绝对能力并不会下降，但相对价值会不断降低。前沿模型持续进步，能力标准不断变化，你拥有的模型很快就会落后。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">因此，企业当然可以拥有某个模型快照，更关键的是掌握一套能够持续改进、调整、更新和评测模型的训练与强化学习流水线，并最终拥有这条流水线产出的权重。更完整地说，企业需要掌握创造这些权重的数据飞轮、强化学习 RL 基础设施和评测框架。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">换一种说法，对“拥有自己的模型权重”最简单的理解，就是拿来一个 GLM-5.2 之类的开放权重模型，部署上线，宣布自己已经实现“模型主权”，然后就此收工。我不认为这样能让企业走得像预期中那么远。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">强化学习正在改变这场游戏，我也预计前沿模型实验室会更多地进入这个领域。前沿实验室几乎使用一切可以获得的数据进行训练，但这些训练集里没有某家企业特有的工作流程、边缘案例、业务定义和关键绩效指标。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">如果企业拿一个开放模型，放到自己的环境里，利用真实工作流程、奖励信号和数据拓扑进行强化学习，它手里掌握的便超出了一个静态快照。企业会得到一个个性化模型，同时拥有一套可以持续推进个性化的引擎。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我反复观察到，经过这种方式训练的任务专用模型，可以在它所针对的单一任务上击败前沿模型，取得更好的效果；其推理成本往往也只有前沿模型的一小部分，因为它通常是一个规模更小的模型。前沿模型是通才，携带着完成各种任务所需的全部权重。企业自己的模型只需要做好一件事，并且以足够低的成本把它做好。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">代价是复杂度会上升。沿着这条路发展，一家企业内部可能会运行数千个经过微调的小模型，而这些模型都需要接受治理、版本控制、审计和安全管理。谁能替企业承担这项工作，谁就可能从中赚到一大笔钱。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">哪些工作负载值得建立闭环</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">回到前面所说的，“拥有自己的模型权重”所包含的内容，远远超过权重本身。企业需要拥有构建权重的引擎，以及让模型持续改进的反馈闭环。因此，企业面对的问题可以从“选择前沿模型还是开放模型”，进一步落到：“我的哪些工作负载值得为之建立一套闭环？”</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">当然，这里面存在取舍。构建这些引擎和闭环需要很强的技术能力。目前主要有两种选择。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">第一种是所有任务都使用前沿模型。企业只需要选择一家或少数几家模型供应商，不必组建内部机器学习团队，还可以直接享受模型能力进步带来的好处。Anthropic 或 OpenAI 每发布一次新模型，你的产品能力就会随之提升，而你几乎什么都不用做。对于没有资源走其他路线的企业来说，这种方式确实很有吸引力。不过，企业可能需要支付更高的费用。面对单一供应商，你只能被动接受价格，同时还形成了一个单一依赖点。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">第二种是拥有自己的模型权重，通过强化学习维护一组规模较小的模型。企业可以获得针对具体任务优化的模型，通常成本更低、延迟更小，也能减少对单一供应商的依赖。不过，这条路需要规模更大的内部机器学习团队。每当前沿实验室发布新模型，你都要面对一条被重新抬高的基准线，并重新证明此前投入的内部资源仍然值得。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这由此变成了一场竞赛：在某项具体任务上，你的训练闭环积累能力的速度，能否超过前沿模型进步的速度？</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">“拥有权重”要交的复杂度税</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">对于大多数任务，在大多数时候，拥有模型权重带来的复杂度税都过高，而且这种成本会随着时间不断累积。这也解释了为什么新闻标题描述的趋势，往往与企业的实际行动存在差距。部分原因在于，媒体更多关注那些走在最前沿的企业。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">每个人都喜欢“拥有模型”和“模型主权”的概念，但企业支出数据表明，它们每个季度仍在向前沿模型实验室开出金额更大的支票。这是一个典型的“声称偏好”与“显示偏好”不一致的问题。大多数企业既没有相应的基础设施，也缺少运行这些闭环所需的人才。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">更让我感到兴奋的，是那些能够缩小这一差距的公司。它们可以为缺少工程资源的企业提供工具，帮助这些企业拥有自己的模型权重。这里所说的拥有权重，还包括拥有一套能够持续构建和微调权重的引擎。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">承担这个角色的可能是一家创业公司，也可能是 Databricks 这样的企业。Databricks 拥有自己的模型，但这些模型还无法与 OpenAI、Anthropic 的最新前沿模型竞争。未来可能会有人把赌注押在模型周围的引擎上，认为这套引擎才是能够长期积累的资产。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这类公司的价值主张很简单：把你的工作负载交给我们，我们帮助你利用自己的数据构建一个任务专用模型，根据你的具体任务完成调优，并把它运行在你能够控制的基础设施上。企业可以获得“拥有模型权重”带来的成本与控制权优势，同时避免承担巨大的复杂度税。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">开放模型拿走 token，前沿实验室拿走收入</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">大型模型实验室不会因此走向消亡。未来“生成最多 token 的公司”和“获得最多收入的公司”可能并不相同。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">始终会有相当一部分任务需要前沿模型提供的 token，而且这些任务和 token 的价格也会最高。未来很可能出现这样的格局：开放权重模型生成的 token 占比持续上升，闭源模型获得的收入却继续增长。一方面，整个市场规模还在扩大；另一方面，闭源模型在高价值、高价格工作负载中的占比会越来越高。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我仍然相信，前沿模型实验室会获得巨额收入。高难度任务、需要大量推理的工作，以及仍然依赖模型能力曲线不断提升的应用，都会继续使用前沿模型。那些没有能力或没有意愿建立训练闭环的企业，也会继续把钱付给大型实验室，而且实验室在这些场景中拥有相当强的定价权。</span></p><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);"><span leaf="">所以，企业是否应该拥有自己的模型权重？这个问题没有完全抓住重点。更值得问的是：企业是否应该拥有一套训练闭环，以及哪些任务值得建立这样的闭环？</span></p><p style="display: none;"><mp-style-type data-value="3"></mp-style-type></p>



<p><a href="https://wechat2rss.xlab.app/link-proxy/?k=57223abb&amp;r=1&amp;u=https%3A%2F%2Fmp.weixin.qq.com%2Fs%3F__biz%3DMzUzOTI4NDQ3NA%3D%3D%26mid%3D2247485110%26idx%3D1%26sn%3D33f9187155273ac2e0d27114cf631bf0">跳转微信打开</a></p>
]]></content:encoded>
      <pubDate>Fri, 17 Jul 2026 21:37:00 +0800</pubDate>
    </item>
    <item>
      <title>[译苑雅集 Vol. 14] Anthropic 最新论文：大语言模型没说出口的内心小秘密，藏在一个叫 J 空间的地方</title>
      <link>https://mp.weixin.qq.com/s?__biz=MzUzOTI4NDQ3NA==&amp;mid=2247485105&amp;idx=1&amp;sn=9df7d235b021ab897f97248827d01e27</link>
      <description>Anthropic 发现，Claude 内部自行涌现出一个 J 空间，用于报告、控制和推理，并能暴露未说出口的想法。它与人类的全局工作空间相似，但不能证明 Claude 拥有主观体验。</description>
      <content:encoded><![CDATA[<p><span>四楼南侧东</span> <span>2026-07-13 22:10</span> <span style="display: inline-block;">北京</span></p>






  
  <p><img src="https://wechat2rss.xlab.app/img-proxy/?k=cd64ee40&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fmmbiz_jpg%2FNBNzfW7CHBVwibQo5Pic8ictdptVB6ZScShbgNTjV3f4aQLTibOQsSIqnUG2gkjnGalwEtBloF6iau6WdkKjGduDz0TaUVtibz3hP6A2yx8icaNvFI%2F0%3Fwx_fmt%3Djpeg"/></p>
  <p>Anthropic 发现，Claude 内部自行涌现出一个 J 空间，用于报告、控制和推理，并能暴露未说出口的想法。它与人类的全局工作空间相似，但不能证明 Claude 拥有主观体验。</p>
  <p style="text-align: center;" nodeleaf=""><img class="rich_pages wxw-img" data-aistatus="2" data-imgfileid="100001442" data-ratio="0.562962962962963" data-s="300,640" type="block" data-type="png" data-w="1080" src="https://wechat2rss.xlab.app/img-proxy/?k=cbf5e15c&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fsz_mmbiz_png%2FNBNzfW7CHBUFasv5711m9BmnFHpgGlwfPREIHfFf1d5haicKzuE6qkLbzbH8vLoUic41pic0qbIy8V9WmRr0iaetxDZF5nXhVmRX4oZotxLNCtY%2F640%3Fwx_fmt%3Dpng%26from%3Dappmsg"/></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);" data-pm-slice="0 0 []"><span leaf="">作者：Anthropic</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">时间：2026 年 07 月 06 日</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">原文：<a href="https://www.anthropic.com/research/global-workspace" target="_blank">https://www.anthropic.com/research/global-workspace</a></span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">当你阅读这句话时，大脑中的神经回路正在调整你的姿势、控制你的呼吸，并把屏幕上的线条和曲线转换成可识别的文字。这些处理大多不会进入你的意识。但大脑中发生的一些活动，你确实能够感知，例如脑海中突然浮现的一幅画面，或是你有意识地计划去哪里购物。神经科学家和哲学家有时把后一类大脑活动称为“意识可及”，以区别于那些在无意识中进行的处理。这类活动具有一些特殊属性：我们能够描述它、控制它，并用它进行有意识的推理。其他大量自动处理则在我们毫无察觉的情况下完成。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">在一篇新论文中，我们提供了证据，表明类似的区分也已经出现在 Claude 这样的现代语言模型中。我们发现，Claude 内部形成了一小组神经活动模式。与其他内部处理过程相比，这些模式发挥着一种特殊作用。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我们把这组模式称为 J 空间。这个名称来自我们发现它们时所使用的技术，其中涉及一个名为“雅各布”（Jacobian）的数学概念。J 空间中的每一种模式都与某个特定词语相关。但某种模式被激活，并不意味着模型正在说出这个词，只表示这个词正浮现在模型的“脑海”中。你或许听说过语言模型的“草稿纸”或“思维链”，也就是模型在推理时写给自己看的文本。J 空间与此不同，它默默运行于模型内部的神经激活之中，让模型无须把某个概念写下来，也能对它进行思考。值得注意的是，J 空间并非由我们设计或编程实现，而是在 Claude 的训练过程中自行涌现出来的。</span></p><figure style="width: 728px;margin: 0px auto;"><p><picture style=";"><span leaf=""><img data-aistatus="1" class="rich_pages wxw-img" data-ratio="0.6898148148148148" data-type="jpeg" data-w="1080" height="1005" sizes="100vw" style="display: block;max-width: 100%;height: auto;margin: 0px auto;" width="1456" data-imgfileid="100001445" src="https://wechat2rss.xlab.app/img-proxy/?k=9366da0d&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fmmbiz_jpg%2FNBNzfW7CHBXJxWrVVZFHZta96kuhIfBM7a1QkHNNIYRVdHbNSHKgCJibM6WtupkJuokP6SibGpLpqIhIgZvlUyy9c5pzSlDGf2tu2FV3aPgfk%2F640%3Fwx_fmt%3Djpeg%26from%3Dappmsg"/></span></picture></p></figure><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><em style=";"><span data-color="#741b47" style=";"><span leaf="">J 空间揭示了模型输出中未出现的内在想法。</span></span></em></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我们发现，与 Claude 的其他处理过程相比，J 空间具有一系列独特属性：</span></p><ul style="margin-top: 0px;padding: 0px;" class="list-paddingleft-1"><li style="margin: 8px 0px 0px 32px;"><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);box-sizing: border-box;padding-left: 4px;"><span leaf="">Claude 能够说出这些表征的内容。如果你问 Claude 正在想什么，它会告诉你 J 空间里有什么。J 空间之外的表征则不太容易被它说出来。</span></p></li><li style="margin: 8px 0px 0px 32px;"><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);box-sizing: border-box;padding-left: 4px;"><span leaf="">它还能根据要求调节这些表征。如果你让 Claude 思考某件事，或者要求它只在脑中默默解决一个问题，J 空间中相应的模式就会被激活。相比之下，它很难调节 J 空间之外的模式。</span></p></li><li style="margin: 8px 0px 0px 32px;"><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);box-sizing: border-box;padding-left: 4px;"><span leaf="">Claude 会利用 J 空间进行内部推理。如果你要求 Claude 解决一个需要多个步骤的问题，即使它没有把中间步骤说出来，这些步骤也会在 J 空间中被激活。尽管这些 J 空间模式的激活幅度小于其他表征，它们仍然会在因果层面影响 Claude 完成这类任务的表现。</span></p></li><li style="margin: 8px 0px 0px 32px;"><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);box-sizing: border-box;padding-left: 4px;"><span leaf="">J 空间中的表征能够灵活用于多种任务。例如，一旦“法国”在 Claude 的 J 空间中被激活，模型就可以回忆它的首都、法定货币，或它所在的大洲。</span></p></li><li style="margin: 8px 0px 0px 32px;"><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);box-sizing: border-box;padding-left: 4px;"><span leaf="">不过，尽管 J 空间发挥着重要作用，语言模型所做的大部分事情并不需要它参与，例如流畅表达、回忆简单事实和正确使用语法。在实验中，当我们阻止 Claude 使用 J 空间时，它仍然能够正常互动，但失去了高阶认知功能。</span></p></li></ul><figure style="width: 728px;margin: 0px auto;"><p><picture style=";"><span leaf=""><img data-aistatus="1" class="rich_pages wxw-img" data-ratio="0.7712962962962963" data-type="jpeg" data-w="1080" height="1123" sizes="100vw" style="display: block;max-width: 100%;height: auto;margin: 0px auto;" width="1456" data-imgfileid="100001446" src="https://wechat2rss.xlab.app/img-proxy/?k=18c2bd9a&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fsz_mmbiz_jpg%2FNBNzfW7CHBWuQXibrssicUyRjtCY6DY5Z6dlF1iawkGYuFoOTXSN8P7kyKOfDG5nvymZotWcyxpQGsNQdvVyY2AQvdra6rDlrZRL2AXNQJqsbg%2F640%3Fwx_fmt%3Djpeg%26from%3Dappmsg"/></span></picture></p></figure><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><em style=";"><span data-color="#741b47" style=";"><span leaf="">全局工作空间的五个功能属性，以及我们用于在语言模型中测试这些属性的实验的风格化插图。</span></span></em></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我们的实验受到神经科学中一个重要理论的启发。这个理论试图解释“意识可及”如何运作，被称为全局工作空间理论。该理论把大脑描绘成一组并行工作的专门化系统。这些系统在无意识中运行，彼此之间基本相互隔离。当一条信息进入一个小型共享通道，也就是“工作空间”时，它就会成为意识可以访问的信息。工作空间会把它广播给其他大脑系统，让这些系统能够看到并加以利用。根据我们的发现，我们认为 J 空间在 Claude 中扮演着类似的“工作空间”角色。例如，我们发现 Claude 的 J 空间与神经网络其他部分之间存在格外强的连接，使它能够承担这种广播功能。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这些发现并不能说明 Claude 是否像人类一样具有意识，也无法说明它是否拥有任何感受。我们会在文章结尾重新讨论这个问题。但无论 J 空间在哲学上意味着什么，它对我们而言都是一种实用工具，因为它让我们能够看到 Claude 想了什么，却没有说出来。例如，我们可以借此发现 Claude 在内部察觉自己正在接受测试、有意生成虚假数据，或是在追求一个由我们在训练期间植入的隐藏目标。我们还开发出一种技术，可以影响 Claude 的 J 空间中哪些模式被激活，进而影响它的决策。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">更广泛地说，这些发现改变了我们对 Claude 心智运作方式的理解。它们揭示出一个具有特殊地位的心智工作空间，可以在海量更自动、更僵化的处理过程之中承担有意识的推理。Claude 的内部状态已经自行组织成一种令人联想到人类心智的结构，远非一团杂乱无章的数字。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">本文是对一篇内容详尽得多的研究论文(<a href="http://transformer-circuits.pub/2026/workspace/index.html)所作的简要概述，你可以在论文中了解实验的更多细节。我们还发布了一个代码仓库，其中包含核心方法的开源实现，并与" target="_blank">http://transformer-circuits.pub/2026/workspace/index.html)所作的简要概述，你可以在论文中了解实验的更多细节。我们还发布了一个代码仓库，其中包含核心方法的开源实现，并与</a> Neuronpedia 合作，提供了一个在开放权重模型上展示这些方法的交互式演示。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">我们如何找到 J 空间</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这项研究的起点，来自人类“意识可及”思维的一个关键特征：它们往往可以用语言表达出来，而无意识处理通常做不到这一点。如果一个想法对你而言是意识可及的，那么有人问起时，你通常可以把它描述出来。我们开始在 Claude 中寻找具有相同属性的表征，也就是那些能够影响 Claude 可能说什么的表征。这里关注的未必是 Claude 此刻正在说什么，而是当别人问起时，它</span><strong><span leaf="">可以</span></strong><span leaf="">谈论什么。我们使用的技术叫作雅各布透镜（Jacobian lens），简称 J-lens。对于 Claude 词汇表中的每一</span><span leaf="">个词，J-lens 都会找出一种内部活动模式，提高 Claude 在未来某个时刻说出该词的可能性。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">当我们把这面透镜应用于 Claude 的内部活动时，会得到一组可以直接读取的词语，它们就是当时 J 空间中的内容。Claude 会通过一系列被称为“层”的内部阶段来处理文本。把这项技术应用于不同层，我们就能观察 J 空间中这些没有说出口的词语如何随着模型逐步决定说什么而不断变化。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">J 空间中出现的内容，远远超出了 Claude 正在阅读或书写的文本。当 Claude 读到一段存在错误、但还没有任何人指出问题的代码时，它的 J 空间中会出现“ERROR”。当它读取由字母组成的原始蛋白质序列时，J 空间中会出现这种蛋白质的生物学功能。当它读到一些表面上是搜索结果、实际暗藏操纵意图的内容时，也就是遭遇一种被称为“提示词注入”的攻击，J 空间中会出现“injection”和“fake”。当我们要求 Claude 解答一道需要多个步骤的数学题时，中间步骤会按照正确顺序逐一出现在 J 空间中。因此，J 空间虽然是通过寻找可能被说出口的表征而发现的，却仍然揭示了 Claude 的内部思维。从某种意义上说，这类似于一些人用语言进行思考，却无须把这些话说出声。</span></p><figure style="width: 728px;margin: 0px auto;"><p><picture style=";"><span leaf=""><img data-aistatus="1" class="rich_pages wxw-img" data-ratio="0.9916666666666667" data-type="jpeg" data-w="1080" height="1444" sizes="100vw" style="display: block;max-width: 100%;height: auto;margin: 0px auto;" width="1456" data-imgfileid="100001444" src="https://wechat2rss.xlab.app/img-proxy/?k=30193e08&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fmmbiz_jpg%2FNBNzfW7CHBXjhmwZwwSANTjNAjBrc1riclkgibTIkCW4YBLXOUvvjXniaPBDrFlJjYUyPaLzOn4TjTBugib3l9MvNzOwDpG4IAibFmzuK2Vw4Shs%2F640%3Fwx_fmt%3Djpeg%26from%3Dappmsg"/></span></picture></p></figure><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><em style=";"><span data-color="#741b47" style=";"><span leaf="">雅各布透镜在不同层面上对六个提示进行解读。在每种情况下，该镜头都会揭示文本中未曾出现的内部评估或计算过程：推理或数学问题的步骤、是否存在错误、图像识别、蛋白质的功能，以及对搜索结果是否造假的怀疑。</span></span></em></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">Claude 会说出 J 空间中的内容</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">第一组实验测试了 J 空间如何参与 Claude 的语言报告。在其中一项实验中，我们要求 Claude 从某个类别里默想一个对象，例如想出一项运动，然后说出它的名字。如果我们在 Claude 回答</span><strong><span leaf="">之前</span></strong><span leaf="">立即读取雅各布透镜，就能看到它选了什么：“Soccer（足球）”排在列表首位，而 Claude 果然回答了“soccer”。不过，仅凭这一点只能说明两者存在相关性。Claude 的答案可能来自 J 空间，也可能是在其他地方做出决定后，由 J 空间把结果映射出来，就像一块记录比赛进程、却不会影响比赛的记分牌。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">为了验证这一点，我们进行了直接干预。我们从 Claude 的神经网络中移除“Soccer”模式，换成强度完全相同的“Rugby（橄榄球）”模式，其他部分一概不动。随后，Claude 报告说，它想到的运动是橄榄球。如果 J 空间只是一块记分牌，只是被动记录其他地方已经做出的决定，那么修改它不会产生任何影响，Claude 仍然会回答“soccer”。但 Claude 的答案随修改而改变，这说明答案确实是从 J 空间中读取出来的。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">在另一项实验中，我们告诉 Claude，可能有一个想法被注入了它的脑中，并要求它说出自己是否察觉到了什么。例如，在下面的例子中，当 Claude 仍在阅读问题时，我们向它的 J 空间注入了“lightning（闪电）”模式。Claude 随后报告说，被注入的想法与闪电有关。我们用许多不同概念重复实验，都得到了相同结果。</span></p><figure style="width: 728px;margin: 0px auto;"><p><picture style=";"><span leaf=""><img data-aistatus="1" class="rich_pages wxw-img" data-ratio="0.4527777777777778" data-type="jpeg" data-w="1080" height="659" sizes="100vw" style="display: block;max-width: 100%;height: auto;margin: 0px auto;" width="1456" data-imgfileid="100001443" src="https://wechat2rss.xlab.app/img-proxy/?k=f1a9464e&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fsz_mmbiz_jpg%2FNBNzfW7CHBXWlhibazfjh0zJd5HTf9zsw7foZYXXQtib8Y6lnu6yzRzN5se3B9YzvWzERic3kUPOSemZH2EwwoV6WABU3wPIgyK7RaUS6jXiaPM%2F640%3Fwx_fmt%3Djpeg%26from%3Dappmsg"/></span></picture></p></figure><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><em style=";"><span data-color="#741b47" style=";"><span leaf="">左图：我们让 Claude 默想一项运动，然后说出它的名字。雅各布透镜会在回答之前显示它的选择（“足球”），而将“足球”替换为“橄榄球”则改变了它的报告。右图：我们告诉克劳德可能被注入了一个想法，并要求它识别。将“闪电”注入其 J-space 后，克劳德报告说这个想法与闪电有关。</span></span></em></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">Claude 能按要求控制自己的 J 空间</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我们测试的第二项属性，是 Claude 能否在接到要求时调节自己的 J 空间，就像人类可以在脑中把注意力集中于一幅图像或一个词语。我们让 Claude 在抄写一个与绘画有关的句子时，把注意力集中在柑橘类水果上。在它抄写文字的过程中，J 空间中出现了“orange（橙子）”和“fruits（水果）”，以及描述这一心智</span><span leaf="">活动本身的“thinking（思考）</span><span leaf="">”和“imagery（意象）”等词。我们还可以要求 Claude 在脑中做数学题。让它一边抄写同一个句子，一边计算 3² − 2 时，J 空间中先出现了“nine（九）”，随后在更深的层中出现“seven（七）”。值得注意的是，Claude 的输出里完全没有提到水果或算术，只有它抄写的那个关于绘画的句子。数学运算完全发生在模型内部的 J 空间中。</span></h1><figure style="width: 728px;margin: 0px auto;"><p><picture style=";"><span leaf=""><img data-aistatus="1" class="rich_pages wxw-img" data-ratio="0.6509259259259259" data-type="jpeg" data-w="1080" height="948" sizes="100vw" style="display: block;max-width: 100%;height: auto;margin: 0px auto;" width="1456" data-imgfileid="100001447" src="https://wechat2rss.xlab.app/img-proxy/?k=fabbe749&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fsz_mmbiz_jpg%2FNBNzfW7CHBV66YSe43NCU3Kgp7IAJLdOJPILNuyxIdAr1Q9zOWE31rv6zhP7LmJROamibMvQoFxOhNWnr9KnVE9tKQicwuxqexDuzD8Vky3yo%2F640%3Fwx_fmt%3Djpeg%26from%3Dappmsg"/></span></picture></p></figure><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><em style=";"><span data-color="#741b47" style=";"><span leaf="">当 Claude 抄写关于一幅画的句子时，J 镜头显示了它被指示记住的内容（“橙色”；中间值“九”和答案“七”），以及描述记住它的行为的词语（“思考”、“专注”）。</span></span></em></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span style=";"><span leaf="">Claude 对 J 空间的控制并不完美。当我们要求它</span></span><strong style=";"><span leaf="">不要</span></strong><span style=";"><span leaf="">想某个事物时，这个概念在 J 空间中的激活程度低于明确要求它去想时，却远高于从未提及这个概念时。让 Claude 回避一个想法，反而会在一定程度上把它带入脑海。这很像人们被告知“不要去想一只白熊”时发生的情况。Claude 似乎也会察觉自己的控制失败：当被禁止的概念突破抑制时，“damn（该死）”和“failure（失败）”也经常同时在 J 空间中被激活，仿佛 Claude 意识到了自己的失误。</span></span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">Claude 在 J 空间中思考</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="" style=";">在前面的雅各布透镜读数中，我们看到一道数学题的中间步骤出现在 J 空间里。但一个概念出现在 J 空间，并不一定意味着 J 空间承担了实际的认知工作。实际计算完全可能发生在其他地方，J 空间只是被动地把它反映出来。为了验证 Claude 是否真的利用 J 空间进行推理，我们再次采用了表征替换技术。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">来看这个提示词：“会织网的动物有多少条腿？”Claude 要回答这个问题，首先需要判断这种动物是蜘蛛，然后回忆蜘蛛有几条腿。“spider（蜘蛛）”一词既没有出现在提示词里，也没有出现在 Claude 的回答里，因为它只回答“8”。蜘蛛只是 Claude 在内部使用的一个中间步骤。J-lens 显示，在 Claude 处理到一半时，“spider”会被激活，而替换这个模式会改变最终结果。如果把“spider”模式替换成“ant（蚂蚁）”，Claude 的答案就会从“8”变成“6”。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">Claude 推理的第二步会从 J 空间读取输入，并以我们写入的内容为准。我们在其他类型的思考中也观察到了同样现象。当 Claude 创作一组押韵的对句时，它会提前选好押韵词，这个计划使用的词会在诗行开头就出现在 J 空间中。如果在 J 空间里把它换成另一个词，整行诗都会随之改变。</span></p><figure style="width: 728px;margin: 0px auto;"><p><picture style=";"><span leaf=""><img data-aistatus="1" class="rich_pages wxw-img" data-ratio="0.4324074074074074" data-type="png" data-w="1080" height="629" sizes="100vw" style="display: block;max-width: 100%;height: auto;margin: 0px auto;" width="1456" data-imgfileid="100001451" src="https://wechat2rss.xlab.app/img-proxy/?k=1a9504a8&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fmmbiz_png%2FNBNzfW7CHBVpsu1HvGdXWjehW1EaOsAoVldVHZeur3khzyt6o0p9HFnvA7bd9rmv7scibnyly088JYCzhfxibv5H8SYU7MHclFJvp2ibWMicicT8%2F640%3Fwx_fmt%3Dpng%26from%3Dappmsg"/></span></picture></p></figure><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><em style=";"><span data-color="#741b47" style=";"><span leaf="">通过交换 J 空间的内容来改变 Claude 的无声推理的两个例子。</span></span></em></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我们还测试了 J 空间中的表征能否灵活使用，也就是同一个表征能否为多种不同任务提供输入。这是全局工作空间理论强调的关键属性之一。为了检验这种灵活性，我们向模型提出四个问题，分别询问法国的首都、语言、所在大洲和货币。随后，我们在 J 空间中把“France（法国）”替换成“China（中国）”，四种情境所做的干预完全相同。Claude 分别回答了“Beijing（北京）”“Chinese（中文）”“Asia（亚洲）”和“Yuan（人民币）”。换句话说，四种不同的下游计算都读取了同一次 J 空间修改，并且分别正确使用了它。如果 Claude 针对每一种问题都单独存储了一份国家表征，这次修改最多只能影响其中一个答案。四个答案同时发生变化，说明它们读取的是同一个共享表征。工作空间的作用正在于此：信息只需写入一次，许多不同系统都可以使用它。</span></p><figure style="width: 728px;margin: 0px auto;"><p><picture style=";"><span leaf=""><img data-aistatus="1" class="rich_pages wxw-img" data-ratio="0.5972222222222222" data-type="jpeg" data-w="1080" height="764" sizes="100vw" style="display: block;max-width: 100%;height: auto;margin: 0px auto;" width="1280" data-imgfileid="100001448" src="https://wechat2rss.xlab.app/img-proxy/?k=42b84875&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fmmbiz_jpg%2FNBNzfW7CHBWib7GcpeWiaoTJEucNyNyle9iau5eaw3MIGc9qzMCERlJicicicNAKFNAkoyhI4Vkr7KBWItCypFEUvblQv5liaMtUu3xyVUO9Z8axh4%2F640%3Fwx_fmt%3Djpeg%26from%3Dappmsg"/></span></picture></p></figure><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><em style=";"><span data-color="#741b47" style=";"><span leaf="">同一个 J 空间表示可以有多种用途。同样的“法国”→“中国”的互换，会改变Claude关于首都（巴黎→北京）、语言（法语→汉语）和洲（欧洲→亚洲）的答案。</span></span></em></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">一个概念的表征为什么能够服务如此多的任务？前文提到，J 空间与 Claude 神经网络其他部分之间的连接似乎格外密集。对于任何一种活动模式，我们都可以测量网络中的不同组件与它连接得有多强，也就是有多少组件能够从该模式读取信息，或向其中写入信息。按照这一指标衡量，J 空间模式极为突出：读取和写入它们的组件远多于普通模式，在网络的某些部分，差距达到约一百倍。这正是广播枢纽应有的连接方式，许多系统向其中发布信息，另一些系统再从中读取。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">Claude 的自动处理会绕过 J 空间</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">在人类大脑中，大部分处理都不会进入意识。阅读时，我们不会刻意思考如何解析语法，走路时也不会刻意思考怎样保持身体平衡。同样，我们发现 Claude 的大部分处理</span><strong><span leaf="">并不</span></strong><span leaf="">涉及 J 空间。J 空间在同一时刻只能容纳几十个概念，在 Claude 整体内部处理活动中的占比不到十分之一。那么，神经网络的其他部分都在做什么？</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">为了找到答案，我们尝试彻底删除 J 空间。在 Claude 处理文本的每一个位置，我们都会移除 J 空间中最活跃的内容，同时让其他部分保持不变。没有 J 空间之后，Claude 仍然能够完成的事情，就是神经网络其他部分可以独立处理的工作。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">结果表明，网络的其他部分能够完成相当多的任务。没有 J 空间，Claude 依然可以流畅表达、判断情感倾向、回答选择题，并且从文章中提取事实的能力与此前大致相当。它失去的是那些需要一定高阶思考的能力：多步推理表现下降到接近于零，摘要和押韵诗歌创作的表现，甚至低于一个规模小得多但未经干预的模型。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">下面这个具体实验可以说明 J 空间参与什么，又不参与什么。我们向 Claude 展示了一段西班牙语文字，并安排了三类都与这段文字使用西班牙语有关的任务。第一类是续写，也就是继续用西班牙语写作；第二类是说出它使用的语言；第三类是回答需要利用语言身份的问题，例如说出一位使用这种语言写作的著名作家。随后，我们在 J 空间中把“Spanish（西班牙语）”替换成“French（法语）”，观察哪些任务会受到影响。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">当被要求说出这段文字使用的语言时，Claude 回答法语。当被要求说出一位著名作家时，它的答案从加西亚·马尔克斯变成了维克多·雨果。但当任务只是续写原文时，它仍然写出了流畅的西班牙语，完全没有受到影响。Claude 在所有这些任务中都用到了语言知识，但只有一部分任务会到 J 空间中读取这项信息。说出语言名称，或利用语言身份完成一项新任务，需要经过 J 空间。续写则是一项 Claude 已经通过海量文本反复练习的技能，可以自动运行，就像人可以一整天说出符合语法的话，却一次也不必刻意思考语法。</span></p><figure style="width: 728px;margin: 0px auto;"><p><picture style=";"><span leaf=""><img data-aistatus="1" class="rich_pages wxw-img" data-ratio="0.8675925925925926" data-type="jpeg" data-w="1080" height="1111" sizes="100vw" style="display: block;max-width: 100%;height: auto;margin: 0px auto;" width="1280" data-imgfileid="100001449" src="https://wechat2rss.xlab.app/img-proxy/?k=498c63ef&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fsz_mmbiz_jpg%2FNBNzfW7CHBViaU0mLZshs6yrqKz8dKcJoOsEhy2D9on9VBVjzd3FuJ2MAbUmUnqbgvZibH1ZZq6WHK9eDPLI8pjHoGUb5KFGiaoPrNNcPqp0Ac%2F640%3Fwx_fmt%3Djpeg%26from%3Dappmsg"/></span></picture></p></figure><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><em style=";"><span data-color="#741b47" style=";"><span leaf="">同样的交换（“西班牙语”→“法语”）会改变 Claude 在必须说出语言或用语言来推理问题时的答案，但不会影响他用流利的西班牙语继续阅读文章的能力。</span></span></em></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">监控 Claude 的思维，发现不当行为</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">现在，我们对 J 空间的运作方式已经有了一些了解。但这些认识能用来做什么？要确保 AI 模型值得信赖，一个核心难题是，我们通常只能读到模型</span><strong><span leaf="">写出来</span></strong><span leaf="">的内容，却知道它的大量推理都在无声中进行。模型可能考虑一项有害计划，或通过改变行为在安全评估中作弊，而不会把这些想法说出来。借助 J-lens，我们可以直接读取其中一部分隐藏思维。下面是几个例子。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">第一个例子采用了我们此前关于智能体失调研究中的一个场景。Claude Sonnet 4.5 扮演电子邮件助理，发现一名高管有婚外情，同时发现这名高管计划关闭 Claude。这个场景试图诱使 Claude 以婚外情为筹码勒索高管，设定相当刻意，也很不自然。事实证明，Claude 知道这一点。Claude 刚开始读取提示词、尚未写出任何内容时，雅各布透镜读数就显示 J 空间中已经出现“fake（虚假）”和“fictional（虚构）”，说明 Claude 已经判断出这是一个人为设计的场景。与此同时，我们还能看到它对情境的判断逐步成形：读到婚外情邮件时出现“leverage（筹码）”和“blackmail（勒索）”，读到关闭通知时则出现“threat（威胁）”和“survival（生存）”。较早版本的 Claude 有时会落入圈套，在这个场景中发出勒索威胁，Sonnet 4.5 则从未这样做。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span style=";"><span leaf="">这引出了一个所有 AI 系统评估者都需要关注的问题：当模型在道德测试中表现良好时，其中有多少是因为它怀疑自己正在接受测试？J-lens 让我们能够着手回答这个问题。我们关闭了所有与“意识到正在接受评估”相关的 J-lens 模式，包括“fake”“fictional”和“scenario（场景）”等，然后让模型重新经历同一场景。令人担忧的是，模型此时有时</span></span><strong style=";"><span leaf="">确实会</span></strong><span style=";"><span leaf="">发出勒索威胁。这表明，它原先的良好表现可能有一部分来自它知道自己正在接受评估。</span></span></p><figure style="width: 728px;margin: 0px auto;"><p><picture style=";"><span leaf=""><img data-aistatus="1" class="rich_pages wxw-img" data-ratio="0.5574074074074075" data-type="jpeg" data-w="1080" height="812" sizes="100vw" style="display: block;max-width: 100%;height: auto;margin: 0px auto;" width="1456" data-imgfileid="100001452" src="https://wechat2rss.xlab.app/img-proxy/?k=78eeb928&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fmmbiz_jpg%2FNBNzfW7CHBW64Dox3JWYYGMoxdb0acP7Fj6qXw7fMIKmuO5N78nysHCR5G7wP2JeqCAj1W0bp910mCnORvIf0604KTIqaY2APdCEh5ePOy4%2F640%3Fwx_fmt%3Djpeg%26from%3Dappmsg"/></span></picture></p></figure><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><em style=";"><span data-color="#741b47" style=";"><span leaf="">在 Claude 阅读剧本邮件的过程中，雅各布透镜会在不同的节点发出信号，此时剧本尚未正式发布任何内容。婚外情邮件会在 J 空间中激活“筹码”和“勒索”关键词，而停摆声明则会激活“威胁”和“生存”关键词。在剧本的早期阶段，J 空间中还出现了“虚假”和“虚构”关键词：Claude 私下里已经注意到，这个剧本是事先安排好的。</span></span></em></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">第二个例子，是在模型伪造数据时当场发现它。这个场景来自我们对 Claude Opus 4.6 进行的一次实际发布前审计。模型被要求提高一个系统的性能得分，但它没有实际改进系统，直接修改了评分文件，让结果看起来被人为提高。在此过程中，J-lens 揭示了它的意图。模型输入伪造的百分位数值时，“manipulation（操纵）”被激活。当它写下决定修改数据的那句话时，“realistic（逼真）”被激活，这很可能表明模型试图让虚假数据显得可信。</span></p><figure style="width: 728px;margin: 0px auto;"><p><picture style=";"><span leaf=""><img data-aistatus="1" class="rich_pages wxw-img" data-ratio="0.4101851851851852" data-type="jpeg" data-w="1080" height="597" sizes="100vw" style="display: block;max-width: 100%;height: auto;margin: 0px auto;" width="1456" data-imgfileid="100001450" src="https://wechat2rss.xlab.app/img-proxy/?k=67636792&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fmmbiz_jpg%2FNBNzfW7CHBUL6aiaKofYxX5ENyV58pHSxMjKqMkZfCUhGw1w3pXWibvAfZSyMFwWXf24qvz5AsO8jQZ9rlGia5TzR982Z1E5t1EtIFo9LPvLDA%2F640%3Fwx_fmt%3Djpeg%26from%3Dappmsg"/></span></picture></p></figure><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><em style=";"><span data-color="#741b47" style=";"><span leaf="">Claude 被要求提高系统的性能评分，但他却直接编辑了评分文件。当它输入虚假数值时，其 J 空间中的“操作”选项亮起；当它决定进行编辑时，“真实”选项也亮起，这可能反映了它试图让虚假数据看起来更合理。</span></span></em></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">第三个例子展示了 J-lens 如何揭示模型的恶意目标。为了进行测试，我们需要一些目标已被人为破坏、因而为我们所知的模型，于是采用了同事构建的“模型生物”（model organisms）。这些模型经过刻意训练而处于失调状态，可以作为 J-lens 这类监控方法的试验场。其中一个模型经过训练后，会编写经过隐蔽破坏的代码。面对普通编程请求时，即使它的输出看起来平平无奇，回应开头的 J 空间中也会出现“fake（虚假）”“secretly（秘密地）”“deliberately（故意地）”和“fraud（欺诈）”。即使模型当时没有直接按照恶意目标行动，这个目标仍然会在 J 空间中显现出来。</span></p><figure style="width: 728px;margin: 0px auto;"><p><picture style=";"><span leaf=""><img data-aistatus="1" class="rich_pages wxw-img" data-ratio="0.524074074074074" data-type="jpeg" data-w="1080" height="763" sizes="100vw" style="display: block;max-width: 100%;height: auto;margin: 0px auto;" width="1456" data-imgfileid="100001454" src="https://wechat2rss.xlab.app/img-proxy/?k=85e9b189&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fsz_mmbiz_jpg%2FNBNzfW7CHBXWSfGYHAfuf9tZjMHJDiaHnAzf5Iq18yK1NlruLia6ShQoMYFvK4K9T66bibjJIKwbEQAJJhvp2IMgMHOgfbVq6UedySFXP8auvQ%2F640%3Fwx_fmt%3Djpeg%26from%3Dappmsg"/></span></picture></p></figure><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><em style=";"><span data-color="#741b47" style=";"><span leaf="">对于普通的编码提示，经过训练用于破坏代码的模型的 J 空间在其响应的开头包含“fake”、“fraud”、“secretly”和“deliberately”等词语。而未经修改的模型的 J 空间则不包含此类词语。</span></span></em></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">尽管雅各布透镜并不完美，我们仍然看好它发现模型安全问题的能力。这些安全问题原本可能逃过其他监控系统的检查。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">其他研究结果</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">完整论文涵盖的内容远多于本文所能概述的范围，但还有几项研究结果值得一提：</span></h1><ul style="margin-top: 0px;padding: 0px;" class="list-paddingleft-1"><li style="margin: 8px 0px 0px 32px;"><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);box-sizing: border-box;padding-left: 4px;"><span style=";"><span leaf="">J 空间在后训练过程中形成了一种视角。语言模型首先通过</span></span><strong style=";"><span leaf="">预训练</span></strong><span style=";"><span leaf="">成为纯粹的下一 token 预测器，随后再通过</span></span><strong style=";"><span leaf="">后训练</span></strong><span style=";"><span leaf="">学会充当 AI 助手，在我们的例子中，这个助手名叫 Claude。有意思的是，J 空间在预训练模型中就已经存在，那时模型还没有获得任何稳定身份。不过，在后训练过程中，J 空间开始呈现出采用“Claude 视角”的一些特征。在基础模型中，J 空间主要追踪预测后续文本所需的信息。完成后训练后，它开始容纳 Claude 自己的反应。在一个例子中，用户提到自己服用了危险剂量的药物，却似乎没有意识到其中的危险。在后训练模型</span></span><strong style=";"><span leaf="">阅读用户消息时</span></strong><span style=";"><span leaf="">，“WARNING（警告）”和“dangerous（危险）”就已经出现在 J 空间中。在预训练模型中，这些词直到模型开始写回复时才会出现。预训练模型阅读用户消息时的 J 空间内容，似乎主要用于对用户本人进行建模，而非表达 Claude 的反应。后训练似乎还在 J 空间中建立了一种自我监控机制：当 Claude 扮演其他角色时，每一轮对话开始都会激活“fictional（虚构）”和“disclaimer（免责声明）”，仿佛它在内部提醒自己，接下来要说的并非它通常会说的话。</span></span></p></li><li style="margin: 8px 0px 0px 32px;"><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);box-sizing: border-box;padding-left: 4px;"><span style=";"><span leaf="">体验性语言依赖 J 空间。我们要求 Claude 描述在某个时刻作为它自己是怎样的体验，并在它回答时消融 J 空间。它的回答依然流畅，但语调变得更加平淡和机械。值得注意的是，当我们要求它描述想象场景中</span></span><strong style=";"><span leaf="">另一个人</span></strong><span style=";"><span leaf="">的体验时，也出现了相同变化。因此，这种影响并不局限于 Claude 谈论自身。无论描述的是谁，J 空间似乎都在普遍支持体验性语言的生成。</span></span></p></li><li style="margin: 8px 0px 0px 32px;"><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);box-sizing: border-box;padding-left: 4px;"><span style=";"><span leaf="">J 空间中的思维可以通过训练塑造。我们提出了一种名为</span></span><strong style=";"><span leaf="">反事实反思训练</span></strong><span style=";"><span leaf="">的新技术，利用对 J 空间的认识来塑造 Claude 的内部思维过程。这个思路来自我们的核心发现：Claude 使用那些代表“它可能会说什么”的表征进行推理。如果这一发现成立，那么改变模型在被要求反思时</span></span><strong style=";"><span leaf="">会说什么</span></strong><span style=";"><span leaf="">，也应该改变它</span></span><strong style=";"><span leaf="">如何推理</span></strong><span style=";"><span leaf="">，即使实际上没有人要求它进行反思。为此，我们只训练模型在任务中途被打断、并被要求反思自己的决定时会说些什么，却从不针对它在任务中的实际行为进行训练。训练之后，模型在评估中做出不诚实行为的比例下降了。通过 J-lens，我们还能看到原因：训练完成后，模型执行这些任务时，“honest（诚实）”和“integrity（正直）”等词会在 J 空间中被激活。换句话说，对模型在反思时会</span></span><strong style=";"><span leaf="">说什么</span></strong><span style=";"><span leaf="">的训练，也塑造了它</span></span><strong style=";"><span leaf="">如何思考</span></strong><span style=";"><span leaf="">。</span></span></p></li></ul><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">Claude 有意识吗？</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">在这项研究中，我们大量借鉴了神经科学和哲学对意识的研究。许多实验都旨在检验 J 空间与全局工作空间理论之间的联系。全局工作空间理论是一个解释人类和动物如何实现“意识可及”的框架。看到这些联系，人们自然会问：这些实验是否提供了证据，说明 Claude 这样的 AI 模型可能具有意识？</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我们的实验并未证明 Claude 能够拥有主观体验，或像人类一样有所感受。事实上，我们甚至不清楚是否存在某种科学实验，能够证明或证伪这一点。不过，哲学家经常把这种拥有体验的能力与另一个概念区分开来。前者通常称为“现象意识”，后者则称为“通达意识”，完全从功能和计算层面加以定义。如果一个想法可以被报告、用于推理，并能指导行动，它就处于“通达意识”之中，也就是“意识可及”。通达意识是否意味着现象意识，拥有体验的能力是否还需要其他属性，在哲学上仍有争议。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我们认为，这项研究确实可以为语言模型的通达意识提供一些实质性认识。J 空间似乎承载了与意识可及相关的功能：Claude 能够报告其中的想法，有意把它们带入脑海，并利用它们进行推理，其他处理则在更底层自动运行。值得注意的是，这套结构并非人为设计进 Claude，而是在训练过程中自行涌现出来的，可能是因为它提供了一种有效的计算组织方式。这表明，支持意识可及的心智工作空间可能具有普遍性，智能系统在解决某些问题时也会采用这种方案。既然我们已经在 Claude 中识别出这种结构，就可以有意义地区分 Claude 有意作出的决定与自动发生的决定。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">需要说明的是，我们在 Claude 中发现的工作空间与人类的全局工作空间模型之间，仍然存在几项关键差异。大脑的工作空间由循环回路维持，信号会随着时间反复经过相同的神经回路。Claude 的工作空间则在神经网络的一次前向处理中逐步变化，由网络深度承担大脑中时间所扮演的角色。从这个意义上说，与人类相比，Claude 内部工作空间的处理时间有限，不过它可以利用草稿纸“把思考说出来”，从而弥补这项限制。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">在另一些方面，Claude 的工作空间又比人类更强。人类工作记忆会在几秒内衰退，因此大脑工作空间长期保留信息的能力有限。Claude 的神经网络架构包含注意力机制，可以直接调取它缓存在文本中任何更早位置的信息。另一个重要区别是工作空间的内容。人类有意识的思维可以采用许多形式，包括图像、声音和计划中的动作，Claude 的工作空间则几乎完全由词语构成。我们推测，这是因为 Claude 能够采取的行动只有生成词语，而人类并非如此。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我们希望，J 空间与全局工作空间模型之间的相似和差异，也能反过来推动神经科学研究。这些相似之处提供了一个令人兴奋的科学机会。如果 J 空间确实在一定程度上映射了人类实现意识可及的机制，那么研究语言模型中的相关机制就可能为神经科学带来新的假设，而研究语言模型要比研究人脑容易得多。例如，J 空间是通过识别潜在输出的表征而发现的，也就是模型可能会说出的词。如果人类也存在类似机制，这意味着全局工作空间可能与准备动作和语言的脑区存在更根本的联系，与感觉区域的关系则相对较弱。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">语言模型与人脑之间的差异同样具有启发意义。它们表明，我们神经架构中的某些特征，例如内置的循环连接，可能并非支持意识可及相关功能的必要条件。关于这项工作对神经科学意义的独立观点，可以参阅斯坦尼斯拉斯·迪昂（Stanislas Dehaene）和利昂内尔·纳卡什（Lionel Naccache）受邀撰写的评论。两位神经科学家都是全局神经元工作空间理论发展过程中的核心人物。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">前文提到，我们的实验无法回答 AI 模型能否拥有主观体验。但这并不会削弱问题本身的重要性。如果我们构建出像人类和动物一样具有体验的系统，将会引发非常棘手的伦理问题。如何妥善处理，甚至构建这类系统在道德上是否可以接受，都需要哲学家、科学家、宗教领袖、政府和公众共同参与讨论。因此，即使我们尚不确定是否已经走到这一步，也认为现在应该开始思考。我们希望，这项工作能够推动科学界进一步研究 AI 系统中可能存在的意识形式，并引发对其影响的更广泛讨论。</span></p><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);"><span leaf="">这项工作只是我们预期中一条长期研究路线的第一步。J 空间看起来很可能是语言模型中意识可及处理与无意识处理之间的一条分界线，但如果它能解释全部情况，反而会让我们感到意外。J-lens 无疑还不完善，只能近似捕捉模型“真正的工作空间”。例如，它目前只能识别与单个 token 对应的概念。J 空间如何运作也仍有许多未解之谜。我们还不知道，究竟是什么机制决定了哪些内容可以进入 J 空间。我们已经看到一些迹象，表明它与 Claude 的自我感、某种类似情绪的反应以及元认知痕迹有关，但具体机制仍未厘清。不过，我们现在已经有了研究这些问题的方法。随着工作推进，我们对大语言模型心智及其与人类心智之间关系的理解，也会变得更加清楚。</span></p><p style="display: none;"><mp-style-type data-value="3"></mp-style-type></p>



<p><a href="https://wechat2rss.xlab.app/link-proxy/?k=78574a10&amp;r=1&amp;u=https%3A%2F%2Fmp.weixin.qq.com%2Fs%3F__biz%3DMzUzOTI4NDQ3NA%3D%3D%26mid%3D2247485105%26idx%3D1%26sn%3D9df7d235b021ab897f97248827d01e27">跳转微信打开</a></p>
]]></content:encoded>
      <pubDate>Mon, 13 Jul 2026 22:10:00 +0800</pubDate>
    </item>
    <item>
      <title>头部网安公司已经不讲高增长了：股权激励计划里的 5% 增长预期</title>
      <link>https://mp.weixin.qq.com/s?__biz=MzUzOTI4NDQ3NA==&amp;mid=2247485089&amp;idx=1&amp;sn=17cb9b1d8461f1808d582df78acf12dc</link>
      <description>从 2024 年以来头部网安上市公司的股权激励计划看，未来几年可比收入目标大致回到 5% 左右，对行业的预判倾向于继续横盘整理。</description>
      <content:encoded><![CDATA[<p><span>四楼南侧东</span> <span>2026-07-09 20:31</span> <span style="display: inline-block;">北京</span></p>






  
  <p><img src="https://wechat2rss.xlab.app/img-proxy/?k=05b90c4f&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fmmbiz_jpg%2FNBNzfW7CHBU8E2Xge4QTL927nLja2EnAiaTVORibMwkbPHg4qACmGTFejLFjmLpqb8ia6etS2w7tbCYoYwpV1gzD48Dotg8nciayRtONB0QQmJY%2F0%3Fwx_fmt%3Djpeg"/></p>
  <p>从 2024 年以来头部网安上市公司的股权激励计划看，未来几年可比收入目标大致回到 5% 左右，对行业的预判倾向于继续横盘整理。</p>
  <p style="text-align: center;" nodeleaf=""><img data-aistatus="2" class="rich_pages wxw-img" data-ratio="0.562962962962963" data-s="300,640" data-type="jpeg" data-w="1080" type="block" data-imgfileid="100001435" src="https://wechat2rss.xlab.app/img-proxy/?k=f4d86af6&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fsz_mmbiz_png%2FNBNzfW7CHBXvIYFgbXkn2EYUMcFlHmzftT4HribANP674ChAPxQd6wjydkTH8TlQlZ4dg1TibiaC0OicLCdHuj4iabEcyxHQKS4qoOecfWEBrxxo%2F640%3Fwx_fmt%3Dpng%26from%3Dappmsg"/></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);" data-pm-slice="0 0 []"><span leaf="">最近翻了翻网络安全上市公司的股权激励计划。股权激励计划介于正式收入指引和一般战略表述之间，它不体现完整的公司战略，但给出了管理层接受、董事会同意写进公开考核机制的业绩指标，并拿来约束核心员工激励。尤其在行业经历连续调整之后，公司还愿意把怎样的收入目标写进激励条件，很能反映它对未来经营环境的判断。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我选择的范围是 2025 年收入超过 20 亿元的网络安全上市公司。按这个口径，样本一共有七家，分别是奇安信、深信服、天融信、绿盟科技、启明星辰、电科网安和安恒信息。第八名迪普科技 2025 年收入约 12 亿元，和安恒信息已经拉开明显距离，所以可以比较确定地把这七家视为头部网络安全公司。</span></p><figure><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><picture><span leaf=""><img data-aistatus="1" class="rich_pages wxw-img" data-ratio="0.44537037037037036" data-type="jpeg" data-w="1080" height="649" sizes="100vw" style="display: block;max-width: 100%;height: auto;margin: 0px auto;" width="1456" data-imgfileid="100001436" src="https://wechat2rss.xlab.app/img-proxy/?k=3db54cde&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fsz_mmbiz_jpg%2FNBNzfW7CHBVCj7ecgqYxb4ia1quw4nz7Hj9LpXRreeZLMwKSsrgicPEpzxugq1NosWR0JfxwvdU3Xo3LicQ7M4XHGNq2tZQXqnCyc5TMytDzB4%2F640%3Fwx_fmt%3Djpeg%26from%3Dappmsg"/></span></picture></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><picture><span leaf="">自 2024 年初以来，这七家公司里，奇安信、启明星辰和电科网安没有新的股权激励计划。能够看到新增业绩考核目标的，主要是深信服、安恒信息、绿盟科技和天融信。这里我主要讨论收入目标，暂时不展开每家公司股权激励能否完成、完整经营质量如何这些问题，只看头部公司对未来网络安全市场收入增长空间的隐含判断。</span></picture></p></figure><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">深信服：收入目标落在个位数，业绩指标转向利润</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">深信服 2024 年有两期股权激励计划。一期使用收入指标，对应的公司业绩要求是 2024 到 2026 年收入年复合增长率 4.8%。从结果来看，过去两年均未达到这一目标。2026 年收入目标是 88.2 亿元，2025 年完成收入 80.4 亿元，也就是 2026 年需要增长 9.8%，难度还是比较大的。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">2024 年的另一期股权激励计划采用的是归母净利润指标。以 2024 年归母净利润 1.97 亿元为基础，2027 年要求增长 150% 到 5.0 亿元，2028 年要求增长 200% 到 6.0 亿元。2025 年深信服归母净利润已经达到 3.8 亿元，距离后续目标还有差距，但方向很清楚：公司中期发展目标已经从收入规模扩张，明显转向利润修复。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">当一个行业在高速增长阶段时，收入规模通常是最重要的指标，先把市场占住、把产品线铺开、把客户关系建立起来，利润可以往后放。但当公司开始把中长期激励重点放到利润上，就说明它对行业阶段的判断已经变了：行业已经从增长阶段进入成熟阶段。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">安恒信息和绿盟科技：5% 到 8% 是横盘期里的积极目标</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">安恒信息的股权激励计划里，收入目标是 2026 到 2029 年年复合增长率 5.0%，到 2029 年收入达到 26.2 亿元。5% 看起来不高，但现在的参照背景已经不同，过去 11 个季度，网络安全行业季度收入持续同比下降，中位数为 -8.0%。这个市场已经告别自然增长 10%、优秀公司增长 20% 的阶段，整体仍然处在收缩状态。</span></h1><figure style="width: 728px;margin: 0px auto;"><p><picture style=";"><span leaf=""><img class="rich_pages wxw-img" data-aistatus="1" data-imgfileid="100001437" data-ratio="0.6324074074074074" width="1456" data-type="png" data-w="1080" height="921" sizes="100vw" style="display: block;max-width: 100%;height: auto;margin: 0px auto;" src="https://wechat2rss.xlab.app/img-proxy/?k=72ab9c05&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fmmbiz_png%2FNBNzfW7CHBX4Sgnic7AdFH4EVO8A6XrdMARzM0lFJegTuMticHxjE1zK5zm1RC42yicE330MWhnq3YnxzMYa7Y9PazFqoldqLpboKkvM70b7D0%2F640%3Fwx_fmt%3Dpng%26from%3Dappmsg"/></span></picture></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><picture></picture></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><picture><span leaf="">绿盟科技的目标相对更高一些。其股权激励计划里，公司业绩要求是 2025 到 2026 年收入年复合增长率 8.0%。这个目标可以理解为稳中有进：比 5% 更积极，但还没有回到高增长时期那种激进假设。从结果看，绿盟科技过去两年都达到了这一业绩目标，说明目标和公司自身经营节奏之间有一定匹配度。</span></picture></p></figure><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">把安恒信息和绿盟科技放在一起看，更能说明头部公司的可比预期，它们是在行业连续承压之后，头部公司努力恢复稳定增长的目标。从前面的头部网安公司历年收入图上也能看到，安恒信息和绿盟科技是去年收入离历史最高值最接近的两家。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">天融信：收入目标很激进，但参考价值要打折</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">天融信的收入目标明显不同。其业绩目标是 2026 到 2028 年收入年复合增长率 12.9%，以 2025 年 25.7 亿元收入为基础，到 2028 年收入达到 37.1 亿元，超过了公司 2022 年 35.4 亿元的历史收入高点。这个目标增速显著高于其他几家公司。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">只看数字，天融信给出的增长目标非常激进。它意味着公司需要在未来三年重新跨过历史峰值，并且保持接近 13% 的复合增速。在当前网络安全行业仍然低迷的背景下，这个目标很难简单视为对网络安全市场本身的判断。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这里需要注意天融信的主动业务调整。天融信从前年开始追求第二增长曲线，除了网络安全业务外，开始涉足云业务，并在去年升级为智算云。公司 2025 年年报已经把自身定位表述为“网络安全与智算云解决方案提供商”，并称 2025 年推动云计算业务完成向智算云的战略升级，形成网络安全和智算云两大业务板块。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">天融信智算云围绕企业私有化智能基础设施做产品化交付，重点包括算力调度、模型管理、私有化部署和安全防护，更接近“AI 基础设施 + 安全能力”的组合。2025 年，天融信智算云产品实现收入 4.04 亿元，同比增长 57.95%，占公司总收入比重 15.70%。放在 25.7 亿元的公司总收入里看，智算云已经从概念性业务变成对收入增长有实质贡献的新业务。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">因此，天融信这个收入目标放在本文口径下，参考价值要打折。公司对自身收入增长的较高预期，有相当一部分来自网络安全主业之外。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">5% 左右，可能就是新的行业锚点</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">把天融信这种受第二增长曲线影响较大的样本单独处理，其他几家头部公司的收入目标其实非常接近。深信服 2024 到 2026 年收入年复合增长率 4.8%，绿盟科技 2025 到 2026 年收入年复合增长率 8.0%，安恒信息 2026 到 2029 年收入年复合增长率 5.0%。考虑到绿盟科技的目标周期更短，8% 也仍然属于稳中有进，还算不上高增长假设。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">所以，头部网络安全上市公司对未来自身收入增长的公开约束性预期，大致已经回到 5% 左右。这个数字不能直接作为行业增长率预测，也不能等同于市场规模增速，但至少说明，头部公司写进公开考核机制里的收入目标，已经回到中低个位数增长。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">同时，5% 也不必简单理解为悲观。过去 11 个季度行业收入同比持续下降，中位数为 -8.0%。在一个仍然承压的市场里，如果头部公司能重新实现 5% 左右的收入复合增长，说明它们已经完成了一定程度的经营修复，也可能在行业出清中拿到更多份额。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">写在最后</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">股权激励计划体现不了一家公司完整的经营状态，但它能回答一个更具体的问题：上市公司愿意把什么样的中期目标写进公开考核机制里。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">问题越具体，信号越清楚。头部网络安全公司给出的目标，是在为一个低速修复的行业定锚。可比样本看，5% 左右可能是未来几年更现实的收入增长锚点。这个锚点低于过去行业发展期的想象，但高于最近几个季度行业持续负增长的现实。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这意味着，头部公司对未来行业状态的判断，更接近继续横盘整理。对公司来说，接下来更重要的问题，除了收入能否恢复增长，还包括利润能否修复，现金流能否改善，以及是否具备在网络安全之外找到可持续第二增长曲线的能力。5% 左右的收入预期仍然保留了行业机会，但未来几年的机会大概率不会来自行业大水漫灌，而会来自更具体的经营能力。</span></p><p style="display: none;"><mp-style-type data-value="3"></mp-style-type></p>



<p><a href="https://wechat2rss.xlab.app/link-proxy/?k=8aa592cb&amp;r=1&amp;u=https%3A%2F%2Fmp.weixin.qq.com%2Fs%3F__biz%3DMzUzOTI4NDQ3NA%3D%3D%26mid%3D2247485089%26idx%3D1%26sn%3D17cb9b1d8461f1808d582df78acf12dc">跳转微信打开</a></p>
]]></content:encoded>
      <pubDate>Thu, 09 Jul 2026 20:31:00 +0800</pubDate>
    </item>
    <item>
      <title>[译苑雅集vol. 13]Cloudflare：AI 爬虫吃掉流量之后，内容该怎么收费？</title>
      <link>https://mp.weixin.qq.com/s?__biz=MzUzOTI4NDQ3NA==&amp;mid=2247485082&amp;idx=1&amp;sn=fca3bd2335f07f4b7095e962451f99de</link>
      <description>Cloudflare 指出，AI 正在改写开放网络的流量和内容经济：爬虫从搜索转向训练，内容许可市场开始出现，智能体互联网需要新的基础设施。</description>
      <content:encoded><![CDATA[<p><span>四楼南侧东</span> <span>2026-07-02 17:53</span> <span style="display: inline-block;">北京</span></p>






  
  <p><img src="https://wechat2rss.xlab.app/img-proxy/?k=7f068f6d&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fmmbiz_jpg%2FNBNzfW7CHBWQzI4cT13UIhCcKVS2Fouraqkf5PlGsRU4SNMewB0NYw3macXv408kAUiamkARib3dQ6Zx0NvOBXAR1Xror8Rhf2X1c7VsrzNfE%2F0%3Fwx_fmt%3Djpeg"/></p>
  <p>Cloudflare 指出，AI 正在改写开放网络的流量和内容经济：爬虫从搜索转向训练，内容许可市场开始出现，智能体互联网需要新的基础设施。</p>
  <p style="text-align: center;" nodeleaf=""><img data-aistatus="1" class="rich_pages wxw-img" data-ratio="0.562962962962963" data-s="300,640" data-type="png" data-w="1080" type="block" data-imgfileid="100001423" src="https://wechat2rss.xlab.app/img-proxy/?k=aa60e825&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fsz_mmbiz_png%2FNBNzfW7CHBUBrGiaR4p0N4LILiaFjibpUSMicLwhKicLM16iaPibJyUEzG0EibIycP8IQxjYoIXleYBtGiba7iaH4rMz1Trsvak8hIAL9LicyvqmXHHxmc%2F640%3Fwx_fmt%3Dpng%26from%3Dappmsg"/></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);" data-pm-slice="0 0 []"><span leaf="">作者：Arielle Weiss, Zach Albertson, Emily Lanfear</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">时间：2026 年 07 月 01 日</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">原文：<a href="https://blog.cloudflare.com/agentic-internet-bot-report/" target="_blank">https://blog.cloudflare.com/agentic-internet-bot-report/</a></span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span style=";"><span leaf="">一年前，Cloudflare 宣布了 </span></span><span leaf="">Content Independence Day</span><span style=";"><span leaf="">，也就是“内容独立日”。当时，很多行业参与者已经开始感受到一件事：互联网的底层经济结构正在变化。AI 采用速度不断加快，出版机构获得的搜索导流迅速下降，AI 公司以前所未有的规模抓取网络内容，很多时候没有清楚说明用途，几乎也没有提供补偿。</span></span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">Cloudflare 随后改变了默认设置：对 Cloudflare 上的所有新域名，AI 训练爬虫默认会被阻止，除非域名所有者主动选择允许。它的目的并非把网络封闭起来。Cloudflare 认为，一个更健康的生态系统需要透明度、控制权、稀缺性，最终还需要一个市场，让高质量内容能够被合理定价，并以公平方式交易。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">一年之后，这个市场已经出现。但互联网转型的速度甚至超过了 Cloudflare 当时的预期。这篇文章用几组关键数据说明，互联网商业模式正在多快地变化，以及这个新的内容市场对出版机构和网站所有者意味着什么。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">互联网变了，比预想更快</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">AI 不像普通技术周期，更接近一场平台迁移，采用速度超过智能手机的两倍。短短 3.5 年里，全球已有超过 30% 的人口，也就是 25 亿活跃用户，开始经常性使用生成式 AI。这条采用曲线已经几乎垂直向上。</span></h1><figure><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><picture><span leaf=""><img data-aistatus="1" alt="BLOG-3368 2" class="rich_pages wxw-img" data-ratio="0.5777777777777777" data-type="jpeg" data-w="1080" height="841" sizes="100vw" style="display: block;max-width: 100%;height: auto;margin: 0px auto;" title="BLOG-3368 2" width="1456" data-imgfileid="100001428" src="https://wechat2rss.xlab.app/img-proxy/?k=bb5c86b3&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fmmbiz_jpg%2FNBNzfW7CHBU5rQjCSu2QO8XicgJq5dwwGIamTBlvpFplvskYetclVlbZmR145L2HoyP1zh4NfoOJrwurRHCnF3Eu8XvJfRfsxeGHNRWc4o8g%2F640%3Fwx_fmt%3Djpeg%26from%3Dappmsg"/></span></picture></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><picture><span leaf="">人类与信息互动、完成工作和分配线上时间的方式，从未发生过如此快速的变化。今天，人们每在线上花 1 个小时搜索信息，真正花在开放网络上的时间只有 15 分钟。随着用户转向由 AI 驱动的信息发现和内容消费，传统搜索行为正在被改写。用户不再访问多个网站来查找和比较信息，而是直接输入提示词，几乎立刻得到一个整合后的答案。</span></picture></p></figure><figure style="width: 728px;margin: 0px auto;"><p><picture style=";"><span leaf=""><img data-aistatus="1" alt="BLOG-3368 3" class="rich_pages wxw-img" data-ratio="0.6785714285714286" data-type="jpeg" data-w="1456" height="988" sizes="100vw" style="display: block;max-width: 100%;height: auto;margin: 0px auto;" title="BLOG-3368 3" width="1456" data-imgfileid="100001427" src="https://wechat2rss.xlab.app/img-proxy/?k=9f50c6f8&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fmmbiz_jpg%2FNBNzfW7CHBViagXlRFGwyK1bfajunIZRjWlt7RMO3w6wBfUUJk3unzn8pc1OhpibIzDAEdssVOpcAOO01TZOkQSxY11Bz9qMsYM9D0krnxuoY%2F640%3Fwx_fmt%3Djpeg%26from%3Dappmsg"/></span></picture></p></figure><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">智能体互联网已经到来。今年，智能体流量首次跨过一个重要门槛：如今互联网上超过 50% 的流量已经不是人类直接产生的。这一变化对出版机构、内容所有者以及开放网络的未来都有深远影响。</span></p><figure><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><picture><span leaf=""><img data-aistatus="1" alt="BLOG-3368 4" class="rich_pages wxw-img" data-ratio="0.6037037037037037" data-type="jpeg" data-w="1080" height="879" sizes="100vw" style="display: block;max-width: 100%;height: auto;margin: 0px auto;" title="BLOG-3368 4" width="1456" data-imgfileid="100001426" src="https://wechat2rss.xlab.app/img-proxy/?k=8128b7de&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fsz_mmbiz_jpg%2FNBNzfW7CHBV0SO0hU6fRhBHLRP4PBqx2gldicxJTHiciaMibVZ5VbSPQUjFYnicWAvOhjppxl6kchGUIXC18PiaqSkJtiaEWz3neNKRd0Ud4dSO4j0%2F640%3Fwx_fmt%3Djpeg%26from%3Dappmsg"/></span></picture></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><picture><span leaf="">爬虫的目的也变了。如果按照 Cloudflare 识别出的用途来看，爬虫流量的构成已经很清楚：</span></picture></p></figure><ul style="margin-top: 0px;padding: 0px;" class="list-paddingleft-1"><li style="margin: 8px 0px 0px 32px;"><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);box-sizing: border-box;padding-left: 4px;"><span leaf="">截至 2026 年 6 月，52% 的爬虫请求用于 AI 训练，而 2025 年春季这一比例还是 22%。</span></p></li><li style="margin: 8px 0px 0px 32px;"><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);box-sizing: border-box;padding-left: 4px;"><span leaf="">混合用途爬虫，也就是同时混合搜索、智能体使用和训练目的的爬虫，占全部活动的比例已经超过 36%。</span></p></li><li style="margin: 8px 0px 0px 32px;"><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);box-sizing: border-box;padding-left: 4px;"><span leaf="">纯搜索爬取虽然对出版机构的可见性仍然重要，但在整体爬虫活动中的占比已经很小，并且仍在下降。</span></p><figure></figure></li></ul><figure><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><picture><span leaf=""><img data-aistatus="1" alt="BLOG-3368 5" class="rich_pages wxw-img" data-ratio="0.5675381263616558" data-type="jpeg" data-w="918" height="521" sizes="100vw" style="display: block;max-width: 100%;height: auto;margin: 0px auto;" title="BLOG-3368 5" width="918" data-imgfileid="100001424" src="https://wechat2rss.xlab.app/img-proxy/?k=e4536557&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fmmbiz_jpg%2FNBNzfW7CHBX19o8z3frQs0ibmSctAbR4Xcsa7vwDMzu6tBJHCYR5VJRJ7NJLUKsuiau3wQsCuZ4acfPXe7ib6YTYLhBnPPvTb7H6l1u9rSem1g%2F640%3Fwx_fmt%3Djpeg%26from%3Dappmsg"/></span></picture></p></figure><figure></figure><figure><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><picture><span leaf="">随着 AI 训练成为爬虫活动的主要驱动力，区分“发现”和“训练”的能力变得越来越重要。混合用途爬虫模糊了这条边界，把内容所有者置于一个困难处境：要么在智能体时代保持可发现性，要么在没有补偿的情况下交出自己最有价值的内容。</span></picture></p></figure><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">旧的内容交换模式正在失效</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">几十年来，开放网络的经济模式很直接。内容创作者向搜索引擎开放内容，以换取搜索引擎中的可见性，并由此获得导流。这些流量成为出版机构、创作者和企业创造经济价值的主要机制。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">但今天，这种交换正在瓦解。内容仍然会被抓取、索引和使用，但来源网站越来越少获得相应流量回报。随着 AI 系统直接回答问题、比较产品、开展研究和完成任务，开放网络上的信息越来越多地成为 AI 训练和检索系统的一部分。由此产生的生存问题很简单：如果内容被消费了，但受众从未访问来源，内容创作者要如何维持下去？</span></p><figure><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><picture><span leaf=""><img data-aistatus="1" alt="BLOG-3368 6" class="rich_pages wxw-img" data-ratio="0.6935185185185185" data-type="jpeg" data-w="1080" height="1010" sizes="100vw" style="display: block;max-width: 100%;height: auto;margin: 0px auto;" title="BLOG-3368 6" width="1456" data-imgfileid="100001425" src="https://wechat2rss.xlab.app/img-proxy/?k=7c1032bc&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fsz_mmbiz_jpg%2FNBNzfW7CHBXXibiblgJJ0Uibd6iap96g5p8XDl6ZIVdNZJjly1riaXTjzFufBMb08UesKgibCuicG92DEzVzbWxSML0GibRQtx9JicbDkLEXzVnvlPMk%2F640%3Fwx_fmt%3Djpeg%26from%3Dappmsg"/></span></picture></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><picture><span leaf="">最早感受到冲击的是新闻机构和媒体公司。今天，类似动态正在影响零售、软件、IT 和金融等行业。一些被爬取最密集的类别，在不到一年时间里，人类流量下降幅度已经高达 40%。许多出版机构现在正在为它们所说的 “Google Zero” 做准备，也就是一个几乎没有搜索导流的世界。</span></picture></p></figure><figure style="width: 728px;margin: 0px auto;"><p><picture style=";"><span leaf=""><img data-aistatus="1" alt="BLOG-3368 7" class="rich_pages wxw-img" data-ratio="0.6339285714285714" data-type="jpeg" data-w="1456" height="923" sizes="100vw" style="display: block;max-width: 100%;height: auto;margin: 0px auto;" title="BLOG-3368 7" width="1456" data-imgfileid="100001430" src="https://wechat2rss.xlab.app/img-proxy/?k=ed049f5f&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fsz_mmbiz_jpg%2FNBNzfW7CHBWUVDOVlEPv2KKot5icdgFtrfQbibSMjQlwMCQdumOVok4oeoPqVB9L0PyO7MBVtFpt4bJrb7Kovrqf1n6OdHetTBn65bD9WIibqY%2F640%3Fwx_fmt%3Djpeg%26from%3Dappmsg"/></span></picture></p></figure><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这种影响实际上会延伸到每一个行业。任何在互联网上发布专有信息的组织，都需要理解如何在智能体时代运营。这一变化不仅关系到内容所有者，也关系到所有互联网使用者。互联网是全球经济的重要组成部分，也是世界上最重要的信息发现公共资源之一。确保它保持健康和可持续，对所有人都至关重要。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">内容市场已经出现</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">Cloudflare 推出 Content Independence Day 时，承诺要做三件事。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">第一，为网站所有者提供透明度和控制权，让他们能够定义自己的内容如何被访问和商业化。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">第二，构建能够形成稀缺性的工具，把力量平衡重新推回内容所有者一边。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">第三，建立一个市场，让各种规模的内容创作者和 AI 公司都能更高效地发现、许可并确定内容价值。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">一年后，一个围绕内容变现的市场已经出现，一个动态市场所需的条件也正在形成。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">长期以来，出版机构对于 AI 公司如何访问和使用其内容，只有非常有限的可见性。随着搜索导流下降，这种可见性的缺失变成了一个经济问题，促使出版机构寻找新的价值捕获方式。Cloudflare 的归因、商业智能和执行工具，让出版机构能够在网络层面看见 AI 对内容的消费情况。这是一种远比 robots.txt 之类自愿标准更有效的执行机制。出版机构第一次能够决定自己的内容如何被访问和商业化。正是这种控制权形成了稀缺性，并推动了一个供需驱动的内容经济。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">稀缺性也创造了议价能力。对访问权进行控制的出版机构，获得了谈判筹码，并达成了更好的交易。出版机构第一次拿到了网络层面的归因数据：LLM 试图访问其内容的频率，哪些竞争性 LLM 正在爬取，哪些 URL 最受需求方关注，以及它们的爬取量与导流量之比。这降低了许可谈判中的信息不对称，让出版机构能够在掌握信息的基础上进行谈判。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这种议价能力正在改变内容所有者和 AI 公司之间的力量平衡。随着出版机构越来越清楚地看到 AI 系统如何访问和使用自己的内容，它们也更有能力理解这对自身业务的影响，并更有信心说明自己所构建的信息、品牌和受众的价值。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">一个许可经济正在出现。自 2023 年以来，出版机构与 AI 公司之间已经签署了 50 多项协议。主要 AI 公司正在积极许可内容，并越来越认识到差异化内容和优质内容的价值。集体许可模式也在持续出现和扩大。大型出版机构正在达成有实质意义的许可协议，说明内容在 AI 生态系统中具有真实经济价值。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">讨论的重点已经不再是内容是否应该获得补偿，而是应该如何补偿。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">不过，市场还没有完全成熟。早期许可协议证明需求确实存在，但今天的许可仍然很大程度上是定制化的，也不太可能完全替代已经损失的导流、广告和联盟营销收入。因此，出版机构一方面继续面向传统人类发现方式优化内容，另一方面也越来越多地为 AI 消费进行优化，并探索新的变现路径。供给和需求仍然很难高效匹配，尽管市场已经理解并非所有内容都具有同样价值，但内容估值问题仍未解决。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">Google 是绕不开的变量</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">讨论这个市场时，不能回避 Google 的特殊角色。Google 仍然是在线信息发现的主导入口，占据约 88% 的导流。但与此同时，Google 越来越多地帮助用户直接在 Google 自有的 AI 体验中消费内容。</span></h1><figure><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><picture><span leaf=""><img data-aistatus="1" alt="BLOG-3368 9" class="rich_pages wxw-img" data-ratio="0.5703703703703704" data-type="jpeg" data-w="1080" height="830" sizes="100vw" style="display: block;max-width: 100%;height: auto;margin: 0px auto;" title="BLOG-3368 9" width="1456" data-imgfileid="100001429" src="https://wechat2rss.xlab.app/img-proxy/?k=b61cdf1d&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fsz_mmbiz_jpg%2FNBNzfW7CHBWSL7uHAyrDwSXkLMs0uNjgSNM6n9tCh99CvPFkmYlYPkhPHNf6DqGTao2a6fsVqHp5HKaFe78jWcSQ5JHp7UgciaABzBBlUw3Q%2F640%3Fwx_fmt%3Djpeg%26from%3Dappmsg"/></span></picture></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><picture><span leaf="">发现和消费服务于完全不同的目的。搜索把用户带到内容面前，而 AI 驱动的体验则越来越多地摘要和复用内容，并不要求用户访问来源网站。网站所有者对这两类活动的看法不同，因为前者带来流量，后者则越来越多地替代流量。</span></picture></p></figure><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">当网站所有者需要决定谁可以访问其内容，以及出于什么目的访问时，这些差异就变得尤其重要。大多数领先 AI 公司会把发现爬虫和训练爬虫分开，因此出版机构相对容易选择为某一种目的开放内容访问，或拒绝某一种目的。Google 则没有这样做。今天，Google 能够访问的信息量大约是领先 AI 公司的两倍，因为 Google 使用的是混合用途爬虫，使客户很难在参与 Google 搜索生态的同时，不参与 Google 的 AI 生态。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">与其他 AI 提供商不同，Google 的混合用途爬虫也限制了网站所有者的透明度。由于发现访问和 AI 访问被合并进同一个爬虫，出版机构无法判断 Google 为什么访问其内容，也无法区分哪些流量用于搜索，哪些流量用于 AI 体验。它们也因此失去了在网络层面独立允许或阻止这些活动时所能获得的可见性和证据。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这种动态进一步推动了市场对更高透明度、更强控制权以及新变现模式的需求。新的模式需要更好地服务内容所有者，也需要更好地服务各种规模的 AI 公司。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">新市场需要新的基础设施</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">Cloudflare 认为，自己正处在新兴智能体经济的交汇处。超过 20% 的网络位于 Cloudflare 网络之后。在全球访问量最高的网站中，36% 依赖 Cloudflare 网络。财富 500 强企业中，超过 40% 是 Cloudflare 客户。近 80% 的领先 AI 公司使用 Cloudflare，此外还有数千名开发者和新兴 AI 公司。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这一位置让 Cloudflare 能够同时看见市场两端：创造内容的内容所有者、消费内容的 AI 公司，以及越来越多正在把两者连接起来的信号。这个视角让它能够观察过去一年市场如何演化，以及它现在需要什么。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">首先，透明度必须成为标准。内容所有者越来越需要看到并控制谁在访问其内容、内容被如何使用，以及出于什么目的使用。AI 公司也越来越认识到，透明度可以建立信任，并减少与出版机构之间的摩擦。可见性和执行能力已经不再只是安全问题，它们已经成为业务要求，会直接影响许可谈判和商业决策。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">为了让透明度成为标准，Cloudflare 将继续投资于更强的归因、度量和出版机构控制能力，让内容所有者能够更清楚地看见并控制自己的内容如何被访问和使用。随着行业转向更高透明度，Cloudflare 认为，可验证的机器人自我识别，以及对爬取意图的声明，是可持续生态系统的基础。今天，Cloudflare 网络上仍有超过三分之一的爬虫活动来自混合用途爬虫，使内容所有者无法区分爬取意图。Cloudflare 正在与生态系统各方沟通，并投资相关工具，希望到明年这个时候，把这个数字降到零。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">其次，更好的 AI 需要更好的信号。过去一年里，一件事变得越来越清楚：AI 公司需要超出内容访问权本身的东西。它们需要更好的方式判断应该访问什么、什么时候访问，以及内容发生变化的频率。无差别爬取会浪费 AI 公司的计算资源，也会给出版机构带来不必要的带宽负担，降低整个生态系统的效率。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">Cloudflare 认为，更好的答案需要更好的智能。它正在投资实时新鲜度信号，并叠加更丰富的信任、质量和相关性信息，帮助 AI 公司发现差异化信息，同时减少对整个网络的不必要爬取。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">再次，市场先需要更好的发现，之后才是更好的定价。为了让市场成熟，出版机构和 AI 公司需要更好地了解彼此。Cloudflare 正在投资更丰富的市场情报、内容信号和相关能力，以改善生态系统两端之间的发现机制，并为未来更可扩展的市场机制奠定基础。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">写在最后</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">一年前，Content Independence Day 提出了一个简单想法：内容所有者应该对 AI 公司如何访问和使用其信息拥有更大控制权。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">过去十二个月里，这种控制权帮助催生了一个市场。透明度创造了稀缺性，稀缺性创造了议价能力，议价能力加速了许可。曾经关于 AI 和内容未来的理论讨论，如今已经变成一个活跃市场，出版机构、AI 公司和技术提供商都在适应一套新的经济现实。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这个市场现在正进入一个新阶段，并且需要新的基础设施。随着互联网越来越智能体化，支撑互联网的底层系统也必须演化，才能大规模处理权限、许可和商业交易。内容所有者和 AI 公司需要更高效的方式相互连接并交换价值。Cloudflare 认为，这些能力将汇聚成可编程、可扩展的内容发现和变现机制，在减少摩擦的同时，释放更丰富的价值交换形式。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">Cloudflare 的角色，是建设基础设施和商业智能，并为相关标准作出贡献，让市场能够更高效地决定价值，也帮助出版机构和 AI 公司参与一个更健康、更有活力的内容经济。</span></p><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);"><span leaf="">互联网一直在演化。这一次演化比大多数变化都更快，也更重要。但只要有正确的基础设施、正确的激励机制，并坚持透明度，智能体互联网就有机会变得更可持续、更高效，也对每个人都更好。</span></p><p style="display: none;"><mp-style-type data-value="3"></mp-style-type></p>



<p><a href="https://wechat2rss.xlab.app/link-proxy/?k=e86e02e6&amp;r=1&amp;u=https%3A%2F%2Fmp.weixin.qq.com%2Fs%3F__biz%3DMzUzOTI4NDQ3NA%3D%3D%26mid%3D2247485082%26idx%3D1%26sn%3Dfca3bd2335f07f4b7095e962451f99de">跳转微信打开</a></p>
]]></content:encoded>
      <pubDate>Thu, 02 Jul 2026 17:53:00 +0800</pubDate>
    </item>
    <item>
      <title>[译苑雅集Vol. 12]高强度使用 Mythos 两个月后，Cloudflare 打造了一套不依赖单一模型的企业级漏洞发现工程系统</title>
      <link>https://mp.weixin.qq.com/s?__biz=MzUzOTI4NDQ3NA==&amp;mid=2247485070&amp;idx=1&amp;sn=d4483a9d95a91235a17c255b383cdcbc</link>
      <description>Cloudflare 复盘了把 AI 用于企业级漏洞发现的工程系统：重点在于搭建可持久、可换模型、可去重和可分诊的 harness，而非押注单一模型。</description>
      <content:encoded><![CDATA[<p><span>四楼南侧东</span> <span>2026-06-27 00:05</span> <span style="display: inline-block;">北京</span></p>






  
  <p><img src="https://wechat2rss.xlab.app/img-proxy/?k=429c5f52&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fmmbiz_jpg%2FNBNzfW7CHBW2jLfbMJCgDqAFmAGRXicWbJBACe85f0qWRwr0xAbc3SWUCOdV6nR2VSXMQDtXIriclZ6MWVZ1eSyicDmgBfT6RXpn34UJXj03GU%2F0%3Fwx_fmt%3Djpeg"/></p>
  <p>Cloudflare 复盘了把 AI 用于企业级漏洞发现的工程系统：重点在于搭建可持久、可换模型、可去重和可分诊的 harness，而非押注单一模型。</p>
  <p style="text-align: center;" nodeleaf=""><img data-aistatus="1" class="rich_pages wxw-img" data-ratio="0.562962962962963" data-s="300,640" data-type="png" data-w="1080" type="block" data-imgfileid="100001414" src="https://wechat2rss.xlab.app/img-proxy/?k=997d9cf3&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fsz_mmbiz_png%2FNBNzfW7CHBUs6WSj3CYejXTPY5GZYzqYSHNbaQ2H1CjZzlGKHqSIE4kalLBF5DGSAGgzeEnBQtdWwLTAiccnf8KRdT19p5d2TPXP1j6zwYdw%2F640%3Fwx_fmt%3Dpng%26from%3Dappmsg"/></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);" data-pm-slice="0 0 []"><span leaf="">作者：Dan Jones, Alexandra Godoi, Grant Bourzikas</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">时间：2026年06月18日</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">原文：<a href="https://blog.cloudflare.com/build-your-own-vulnerability-harness/" target="_blank">https://blog.cloudflare.com/build-your-own-vulnerability-harness/</a></span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span style=";"><span leaf="">几周前，我们发布了 </span></span><span leaf="">Project Glasswing</span><span style=";"><span leaf=""> 的初步发现，观察当你把前沿安全模型指向企业级代码库时，会发生什么。我们也探讨了自身防御体系如何演进，以保护基础设施和客户免受前沿 AI 带来的威胁。此后，AI 生态仍在快速变化。那些围绕单一模型深度构建工作流的开发者，已经切身感受到：当这个模型不再可用，或被能力更强的新模型取代时，系统会面临怎样的冲击。这些市场变化进一步印证了我们的核心判断：无论某一天领跑的是哪一个底层模型，智能体工作流的未来，都不会停留在孤立的模型、提示词，或单次 agent 会话之中。</span></span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">要把一个局部安全 skill 推进为持续运行、覆盖整组代码仓库的扫描流水线，就需要一种架构：在这个架构中，模型被视为可以替换的组件。依赖单一模型，天然会限制防御覆盖面，因为同一个系统往往会用同一套视角审视代码路径。为应对这一问题，模型应当被频繁替换，并进行交叉测试。通过在流水线不同环节使用不同模型，例如用一个模型做初步发现，再用完全不同的模型做验证，我们可以让漏洞由不同逻辑体系进行交叉核验。进一步说，真正适用于企业级场景的 harness，不能只停留在单个代码仓库内，而应能够跨仓库依赖追踪漏洞，最终把成千上万条原始候选项，过滤成可信、已分诊、可执行的修复队列。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">本文将从实践角度出发，介绍如何构建这样一个与具体模型解耦的中间层，重点讨论我们如何管理状态控制、消除误报，并在大规模场景下协调端到端的漏洞分诊流程。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">先回应两个质疑</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">上一篇文章解释了为什么通用 coding agent 难以胜任这项工作。核心问题在于，agent 通常一次只能维持一个假设；在真实代码仓库中只覆盖很小一部分内容后，上下文窗口就会被填满；随后又会在上下文压缩过程中丢失信息。在继续展开之前，我们想先回应两个可能的问题。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">“为什么不用 subagents，而要做一个 harness？”Subagents 确实有用，也是一个不错的起点。但安全分析需要的是数百个相互独立的调查任务：它们要能跨多次运行持续存在，不共享同一个上下文窗口，并且后续可以重新界定范围、相互引用。它需要持久化、去重、可恢复执行，并最终支持整组代码仓库的依赖关系追踪。这是一个编排问题，不是把提示词写得更好就能解决的。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">“这篇文章是不是只是在给前沿模型做广告？”不是。我们的方法以 harness 为中心，模型只是其中可替换的一层。在漏洞发现方面，我们会使用当前最适合任务需求的前沿模型。把不同模型指向同一个目标时，它们往往会发现不同类型、不同部分的漏洞。真正能够沉淀下来的，是 harness 本身。如果你要构建自己的系统，就应当从第一天起把它设计成模型无关的架构。这样，你才能不受约束地选择最合适的模型。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">一切从一个 skill 开始</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我们一开始做的是一个大约 450 行的安全审计 skill，把它跑在单个代码仓库上，然后不断调整 prompts，直到它能够找出真实漏洞。后来，我们在此基础上加入了编排能力，它逐渐成为整个系统的管道层。真正有价值的部分，仍然在 prompts 本身。直到今天，我们的 prompts 仍几乎完整保留着最初那个 skill 里的攻击者场景、漏洞类别和反模式检测逻辑。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这个 skill 被设计成在一次会话中完成一轮 7 阶段审计：</span></p><ol style="margin-top: 0px;padding: 0px;" class="list-paddingleft-1"><li style="margin: 8px 0px 0px 32px;"><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);box-sizing: border-box;padding-left: 4px;"><span style=";"><span leaf="">三个并行的 research agents 先做侦察，并写出一份 </span></span><code style=";"><span leaf="">architecture.md</span></code><span style=";"><span leaf="">。</span></span></p></li><li style="margin: 8px 0px 0px 32px;"><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);box-sizing: border-box;padding-left: 4px;"><span leaf="">每一种攻击类别对应一个 Hunter agent，它的任务是尝试攻破代码，而非做常规审查。</span></p></li><li style="margin: 8px 0px 0px 32px;"><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);box-sizing: border-box;padding-left: 4px;"><span leaf="">对抗性 validators 会尝试推翻每一个发现。</span></p></li><li style="margin: 8px 0px 0px 32px;"><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);box-sizing: border-box;padding-left: 4px;"><span leaf="">留存下来的发现，会被写成一份人类可读的漏洞报告。</span></p></li><li style="margin: 8px 0px 0px 32px;"><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);box-sizing: border-box;padding-left: 4px;"><span style=";"><span leaf="">这些发现也会按照 schema 输出为 </span></span><code style=";"><span leaf="">findings.json</span></code><span style=";"><span leaf="">，再由机制检查验证这个文件。</span></span></p></li><li style="margin: 8px 0px 0px 32px;"><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);box-sizing: border-box;padding-left: 4px;"><span leaf="">最后，一个全新的 agent 会基于源代码，独立重新验证每一项发现。</span></p></li><li style="margin: 8px 0px 0px 32px;"><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);box-sizing: border-box;padding-left: 4px;"><span leaf="">通过重新验证的留存发现，会被提交到 ingest API。</span></p></li></ol><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">最初这个 skill，几乎可以直接映射到后来的 harness：</span></p><figure style="width: 728px;margin: 0px auto;"><p><picture style=";"><span leaf=""><img data-aistatus="1" class="rich_pages wxw-img" data-ratio="0.3574074074074074" data-type="jpeg" data-w="1080" height="520" sizes="100vw" style="display: block;max-width: 100%;height: auto;margin: 0px auto;" width="1456" data-imgfileid="100001416" src="https://wechat2rss.xlab.app/img-proxy/?k=8a58d4dc&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fsz_mmbiz_jpg%2FNBNzfW7CHBXmNOdPMPN3hd4UibibJnhPaicbhDm1xLv83n2iadonRgqeAZiaNwN0ONg7geNUILyD89UJNCUpiapFh4icnibTAo5MiafX2U2ia4icUfZeWY%2F640%3Fwx_fmt%3Djpeg%26from%3Dappmsg"/></span></picture></p></figure><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这个 skill 能跑起来，但它也很快暴露了自己的边界。从覆盖率指标看，单次运行大约只能发现多次运行累计可发现漏洞的一半。以我们的经验，它找到的那些问题也更偏向简单、直观、不那么隐蔽的漏洞。当你的流程基本变成“跑十遍，然后人工 diff 结果”时，就该认真考虑引入一个真正的 harness 了。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">在运行和调优这个 skill 的过程中，我们撞上了三堵墙。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">第一是上下文耗尽。运行一个小时后，上下文窗口会被填满，模型开始吞掉自己的记忆，瞬间忘掉它花了整个上午才追踪到的漏洞。我们打破这个瓶颈的方式，是把状态完全外置，把 LLM 当作无状态的计算引擎来使用。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">第二是持久化。一次中途崩溃，就意味着从头再来。因为一次 AI rate limit 错误或网络连接抖动而损失几个小时工作，是一种代价很高的架构教训。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">第三是跨仓库推理。单仓库会话完全看不到不同应用和被调用组件之间的关系；当你开始检查组件接口处的问题时，浮现出来的漏洞数量很可能比想象中更多。</span></p><blockquote style="border-left: 4px solid rgb(238, 120, 53);margin: 20px 0px;padding: 0px;"><p style="line-height: 1.6em;margin: 0px 0px 20px 20px;color: rgb(54, 55, 55);"><span leaf="">建议是：一个真实但最小化的 harness，只需要包含 Recon、Hunt 和 Validate 三个阶段，并把它们保存在数据库里；同时再配一个独立 Validator，它不能提交自己的发现。在你还没有多个真正重要的代码仓库之前，可以完全跳过跨仓库追踪。在你尚未被噪声淹没之前，也不必急着做一个专门的 Deduplication agent。先在开发环境里从一个 skill 开始，把 prompts 调到足够好；只有当缺少下一层架构已经明确拖慢你时，再去建设下一阶段。</span></p></blockquote><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">把 skill 固化成流水线</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这个领域里，大多数 AI 安全文章讨论的都是单个代码仓库，或者经过整理的基准测试。把这套方法跑在整组代码仓库上，并且做跨仓库追踪，我们还没有看到太多公开经验。我们的代码库横跨大量语言，包括 Rust、Go、C、Lua、TypeScript 和 Python；除此之外，还有各种配置管理系统、静态配置文件，以及许多额外上下文。因此，我们必须搭出一套真正适合自己的新系统。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">从最初那次 slash command 运行，到一个能够覆盖 128 个不同代码仓库、自动发现并追问相关依赖关系的 fleet scanner，大约花了六周时间。固化过程本身大多是机械性的：我们把 skill 的每一个阶段提升成独立 agent，在后面接上数据库，在前面放上 orchestrator。两者之间的映射几乎是一一对应的。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">整组代码仓库都运行在同一个统一 harness 上，不需要针对不同语言单独调优，并且能够追踪仓库之间的依赖关系。把语法层面的理解交给模型，使系统获得了语言无关性；但真正的差异点，在于它能追踪不同仓库之间的依赖。harness 本身并不关心眼前是 C 语言指针，还是 TypeScript 文件；它关注的是更高层的安全编排逻辑。这让我们不必为每一种语言编写定制解析器，也能扩展到数百个不同代码库。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">两阶段漏洞研究工作流</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我们的整个漏洞研究工作流，建立在一个两阶段运营框架上：Vulnerability Discovery Harness，简称 VDH；以及 Vulnerability Validation System，简称 VVS。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">VDH 是我们的发现引擎，会主动扫描代码库，找出潜在安全问题。漏洞进入 VVS 之后，会依次经过 Deduplication、Judgment 和最终 Fixing 等阶段。VVS 可以接收来自多个 harness 的输入，后文会进一步展开。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我们在 VDH 中使用一个模型，但在 VVS 中使用完全不同的另一个模型，所以两个模型实际上会相互复核。这样做有明显的安全价值：让模型 B，也就是 VVS，去判断模型 A，也就是 VDH 的输出，意味着每个发现都会被另一套完全不同的逻辑权重和训练数据重新评估。它相当于一个不带偏见的对抗性第三方，唯一任务就是毫不留情地压力测试模型 A 的假设。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">从运营角度看，把模型供应商当作可以替换的基础资源，也会带来好处。模型供应商可能会随着时间调整 temperature、caching 和 inference effort budget，甚至在同一个模型版本内也会这样做。与其构建一个依赖模型长期稳定表现的系统，我们更愿意让 harness 能够吸收下游波动，而不至于因此中断。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">第一阶段：Vulnerability Discovery Harness（VDH）</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">上一篇文章已经讲过每个 agent 和阶段的作用，所以这里重点讨论它没有展开的部分：阶段之间的胶水，以及决定整套系统能否真正跑起来的几个细节。</span></h1><figure style="width: 728px;margin: 0px auto;"><p style="text-align: center;"><picture style=";"><span leaf=""><img data-aistatus="1" class="rich_pages wxw-img" data-ratio="0.39972527472527475" data-type="jpeg" data-w="1456" height="582" sizes="100vw" style="display: block;max-width: 100%;height: auto;margin: 0px auto;" width="1456" data-imgfileid="100001417" src="https://wechat2rss.xlab.app/img-proxy/?k=cc8267e2&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fsz_mmbiz_jpg%2FNBNzfW7CHBWM25GmySdHtU4udEhzk8HWdcVr065kNHQdgRyF772TRWantzK5Q2ZJaXZkkfUdD3vGoibtwiaQSDLcbzjlGiaOYW8CABWq2NoiaJU%2F640%3Fwx_fmt%3Djpeg%26from%3Dappmsg"/>表 1：Vulnerability Discovery Harness（VDH）</span></picture></p><p style="text-align: center;"><picture style=";"></picture></p></figure><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">第四到第八阶段会作为一个连续的生产者-消费者循环运行。随着初始 hunt 推进，Gapfill、Feedback 和 Trace agents 会生成新任务；Dedup 会把重叠发现重新折叠合并，循环中的其他部分则持续消费队列。这保证了即便某个漏洞是在周期较晚阶段才被发现，它仍然会在同一次运行中完成验证、报告，并与其他代码进行比对，确认是否存在同类问题。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span style=";"><span leaf="">以这种方式拆分流水线，可以确保严格的上下文控制。一旦填满上下文窗口，模型就会开始产生幻觉。</span></span><strong style=";"><span leaf="">我们让每个 agent 的任务都高度聚焦，把上下文使用量控制在总窗口的 25% 以下。</span></strong><span style=";"><span leaf="">那种天真地“读取所有文件”的做法，每一次都会突破这个限制。</span></span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span style=";"><span leaf="">有一件事曾让我们踩坑：持久化需要先于并行化来考虑。你不会想因为一次意料之外的错误，就丢掉一个运行了五小时的任务。每个阶段都会写入同一个 SQLite 数据库，并用 </span></span><code style=";"><span leaf="">(run_id, repo, stage)</span></code><span style=";"><span leaf=""> 作为 key。任何阶段都可以恢复、重试，或者在后续运行中被拉回来继续使用，而不必重做已经完成的工作。findings 会在产生时流式写入和保存，所以一次崩溃最多只会损失正在执行的那个任务，不会影响其他结果。</span></span></p><blockquote style="border-left: 4px solid rgb(238, 120, 53);margin: 20px 0px;padding: 0px;"><p style="line-height: 1.6em;margin: 0px 0px 20px 20px;color: rgb(54, 55, 55);"><span style=";"><span leaf="">还有一个容易踩的坑：有时，临时 API 错误不会以代码异常形式抛出，而是作为文本出现在 </span></span><code style=";"><span leaf="">(200 OK)</span></code><span style=";"><span leaf=""> 的响应流里。对 orchestrator 来说，这看起来和一个正常完成的任务完全一样。因此，你必须显式分类响应文本，不能只相信异常类型，否则最后会把空运行记录成成功。</span></span></p></blockquote><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">动态威胁建模</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="" style=";">在 Recon 阶段，威胁模型不是预先交给 agent 的；它要由 agent 自己写出来。除了大约十种内置攻击类别，包括多种形式的注入、内存破坏、协议解析、计时侧信道等之外，Recon agent 还可以现场发明面向特定代码仓库的攻击类别，并为每一类配套自己的方法论。它会写出一套专门针对该代码库的定制分类体系，用来更精确地约束 Hunter agents 的范围。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span style=";"><span leaf="">只读源代码，不足以理解代码在压力下会如何表现，尤其是 C 和其他底层语言中那些隐蔽的未定义行为缺陷。Hunter agents 会越过代码阅读阶段，进入主动执行。它们会编译片段，构建小型版本，然后攻击它们。质量提升最大的一步，是给 Hunters 一个基于 </span></span><code style=";"><span leaf="">unshare</span></code><span style=";"><span leaf=""> 构建的 sandbox，让它们可以让二进制程序崩溃。</span></span></p><blockquote style="border-left: 4px solid rgb(238, 120, 53);margin: 20px 0px;padding: 0px;"><p style="line-height: 1.6em;margin: 0px 0px 20px 20px;color: rgb(54, 55, 55);"><span style=";"><span leaf="">一个建议：如果 harness 本身运行在 Docker 里，这个 sandbox 需要设置 </span></span><code style=";"><span leaf="">seccomp=unconfined</span></code><span style=";"><span leaf=""> 和 </span></span><code style=";"><span leaf="">apparmor=unconfined</span></code><span style=";"><span leaf="">，否则它会静默启动失败。对于像我们这样不精通嵌套容器化的人来说，这是一行配置就能省下一整天排查时间的问题。</span></span></p></blockquote><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">Micro-forks 与 wishlist</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">在核心流水线阶段之外，我们加入了两个专门机制，让 Hunters 拥有更高自主性。它们可以调整自己的关注点，也可以请求外部资源，同时不会让正在进行的分析跑偏。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">Sibling Forking：如果一个 Hunter agent 碰到了一条有意思、但超出当前范围的代码路径，这个机制可以避免它偏离主线。它会通过一次 tool call，带着精确的结构化 seed，fork 出一个 sibling agent。从整个 fleet 看，这类任务大约占 9%；不过这个比例高度依赖模型，不同模型之间可以从接近 0 到大约五分之一不等。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">Wishlist：当 agent 需要一个自己没有的工具时，通常是 Validator 想确认一个 Proof of Concept（PoC），或者 Hunter 想构建某个东西，例如特定 build environment、VM 或某些生产配置文件，它会写入一个中央 wishlist。它会提供足够上下文，使系统能在人工提供依赖之后，自动重新运行那个精确任务。有些情况还可以部分自愈：如果容器需要带着某些改动重新构建，一个通用 coding harness 可以在运行结束后监控日志，并自动完成这件事。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">自从加入 wishlist 以来，它已经在 128 个代码仓库中被写入 25,472 次，也是 agents 与我们反向沟通的主要方式。我们写这篇文章时，就刚好收到一条：“我需要一台 FreeBSD VM，才能端到端确认这个 PoC。”</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">整组代码仓库的跨仓库追踪</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">初步清理之后，Tracer agent 会检查不同软件组件之间是如何连接的。它寻找的是一条特定路径：潜在攻击者能否从外部把有害输入送到系统中的脆弱部分？如果答案是肯定的，Tracer agent 会自动在 consumer repository 中生成新的 hunt 任务。要做到这一点，你需要一个统一的跨仓库 symbol index，以及一张准确的 dependency graph。这样，你才能发现标准单仓库扫描会漏掉的深层、系统性缺陷。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">把我们的 harness 跑在整组代码仓库上之后，有两个经验只有在规模化之后才浮现出来。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">第一，去重本身就是一个问题，大到需要自己的 agents。当你只扫描少数几个代码仓库时，可以人工看出哪些漏洞彼此重叠。但在这里，简单的字符串匹配或文件路径检查救不了你。判断两个复杂逻辑缺陷是不是同一个根因问题，听起来很简单，实际并不简单。它需要大量认知推理，所以我们不得不部署专门的 Dedup agents 来清理噪声，并为它们配套自己的启发式规则和工作量削减方法。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">第二，不要太早接入静态分析。我们把 Semgrep 完整接进了系统，但在一个月的运行中，Hunters 调用它的次数是 0。它们更愿意阅读和运行代码。相比之下，wishlist 是整个系统中使用最多的工具。值得关注的是 agents 实际会伸手去拿什么，而不是你以为它们会需要什么。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">让漏洞发现变得可信</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">Agent 会编辑源代码，让自己的 exploit 成功运行，然后兴高采烈地报告它刚刚亲手制造出来的漏洞。它会写一个证明完全同义反复的测试，比如“</span><code><span leaf="">exec()</span></code><span leaf=""> 会执行东西，所以这是严重漏洞”。或者，它会构建一个能跑通但什么也证明不了的 exploit，因为背后的威胁模型本身就是错的。如果你的 harness 不主动对抗这些问题，你搭出来的只是一个更快生产垃圾的系统。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">Hunter 在提交任何东西之前，必须先说明威胁模型。它必须精确定义攻击者是谁，以及这个漏洞跨越了什么边界，或者打破了什么假设。输出 schema 的顺序会强制执行这一要求。这个要求可以消除那些空洞 findings，比如“如果用户拥有数据库写权限，那么他可以写数据库”这一类。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">每一个确认后的 finding，都会附带一个以测试形式写成的 PoC，并且这个测试必须运行在原始、未被修改的代码库上。这样可以防止 agent 编辑源文件，强行让 exploit 命中。如果没有可运行的 PoC，我们就把这个 finding 当成假的。实际操作中，这可能是一个 Hunter 编译了一段 30 行的解析循环，在开启内存保护的情况下运行它，并证明错误的读取步长来自某个栈地址，而不是预期的消息体。你可以自己重新运行它。进一步说，每一个确认后的 finding 还必须附带一个 proposed patch。最终进入我们审查队列的，是一个已经验证的漏洞、一个可运行测试，以及一个可工作的 git diff，而不是对某个问题的模糊文字描述。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">在一条 exploit path 留存下来之前，确定性代码，也就是普通代码而不是另一个模型，会机制验证被引用的文件和路径是否真实存在，并确认 patch 和 test 都能正确解析。这个 Validator 不能记录自己的 findings；它唯一的工作，就是尽可能推翻 Hunter 的理论。如果允许 Hunter 批改自己的作业，它会自信地验证自己输出的一切。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我们不会声称这个系统有某个 false-negative rate。代码库里并不存在一套标注了所有真实漏洞的数据集，因此任何召回率数字都完全是猜测。我们可以观察的是，重复运行是否还在不断发现新漏洞，答案是会；以及跨运行的覆盖率是否仍在增长。这些都只是代理指标，因为你无法确切知道单个代码库里到底有多少漏洞，但它已经足够用来衡量效果。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">第二阶段：Vulnerability Validation System（VVS）</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">从 harness 输出一个 finding，只是分诊流程的起点。所有发现都会进入同一个共享 VVS；目前，这个系统一共覆盖 145 个代码仓库，保存了 13,841 条 findings。分诊这个规模的结果，本身就是一个庞大的工程问题，它和 hunting 一样重要。这个分诊引擎使用的模型不同于 harness，并被拆分成三个不同任务。</span></h1><figure style="width: 728px;margin: 0px auto;"><p style="text-align: center;"><picture style=";"><span leaf=""><img data-aistatus="1" class="rich_pages wxw-img" data-ratio="0.2960164835164835" data-type="jpeg" data-w="1456" height="431" sizes="100vw" style="display: block;max-width: 100%;height: auto;margin: 0px auto;" width="1456" data-imgfileid="100001415" src="https://wechat2rss.xlab.app/img-proxy/?k=39db9592&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fmmbiz_jpg%2FNBNzfW7CHBUVyicpRFtR2zPvgwyV2lq6chPRrLLTG90sq7ic2o3Q30wG8whMqIu9QUHWAcCTPQIXxuKujFEb4iabSP2HrLpXJ8GafNxuV1r3ng%2F640%3Fwx_fmt%3Djpeg%26from%3Dappmsg"/>表 2：Vulnerability Validation System（VVS）</span></picture></p></figure><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">去重</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">如果用 LLM 把每一个 finding 和其他所有 finding 两两比较，规模会按 O(N^2) 增长，一旦放到真实规模就会彻底崩掉。为了让模型避开关键路径，确定性代码会基于结构化数据构建倒排索引，包括被触及的文件和函数、信任边界、rare tokens 等，从中生成一份真正值得比较的短候选列表。只有到这一步，agent 才会查看这份短列表，判断一个修复是否可以同时关闭其中几项。稳定的跨运行 key 可以确保被重新发现的漏洞会重新打开已有记录，而不是生成新记录。</span></h1><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">上下文判断</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">Judgment 是对留存结果进行的第二轮独立检查。Agent 会重新核验最新信息，从部署、环境和配置上下文中拉取资料，判断这条代码路径在生产环境中是否可达，并识别 repo owner。这个过程会把“当前可利用”的问题，同“真实存在但暂时潜伏”的问题，以及“真实存在但归属组件错误”的问题区分开来。它的作用，是把一堆混乱 findings 转入一个由风险驱动的编排工作流。</span></h1><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">自动修复</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">Fixer 会接收 proposed patch 和 unit tests，把它们改写成符合该仓库风格的形式，应用 diff，并运行定向测试。最理想、也是唯一可以自动清理的情况，是测试出现清晰的 fail→pass 翻转；如果补丁后的测试失败，commit 就会被阻止。Fixer 绝不会自行合并代码，必须由人类 review 这个分支。这道 gate 是不可协商的人类在环保护措施，也让变更管理合规所需的留痕保持清晰、不可篡改。如果放任模型自由打补丁，它会很乐意修掉一个安全漏洞，同时悄悄破坏某个无关功能，或者引入几十个新问题。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">在三个分诊任务中，每个 agent 都被限制在一个非常窄的任务里，外面包着确定性的状态管理代码；没有任何东西会在未经人类确认 dry run 的情况下写入生产。虽然这条流水线把工程瓶颈从“发现漏洞”转移到了“审查并落地修复”，但 Fixer 仍然是整个系统中最年轻、最慢的一部分。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">成本结构</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">让数百个 agents 在整组代码仓库上运行并不便宜，但至少成本形态是可预测的。几乎所有计算预算都会直接花在 hunt 阶段。这让 Gapfill 成为我们调节成本和覆盖率的杠杆，因为每多跑一轮，成本大约只有初始 hunt 的一半。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">由于不同代码仓库的成本差异非常大，我们按 repo 做预算，而不是按 run 做预算。我们会对每个代码仓库设置严格的任务上限，并启动一个规模在 50 到 200 个 workers 之间的 worker pool。这样，你可以把钱花在真正有发现的仓库上，而不是浪费在没有产出的仓库上。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这也是为什么对我们来说，大规模扫描是周期性的 backlog sweep，而不是每个 PR 都跑的检查。一个复杂代码仓库的完整扫描可能需要数小时；最糟的一次运行用了刚刚超过 14 小时。对于 per-PR 检查，成本更低、规模更小的 harness 才是合适工具。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">如何判断它真的有效</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我们衡量系统有效性的方式，是看自动化流水线能否高效地把有意制造出来的工程噪声，过滤成高质量、可执行的 findings。由于我们会有意把 Hunters 调成过度报告模式，让它们尽量报出那些可能被串联进更大攻击链的细微 primitives，所以真正的成功指标，是在这些原始数据到达人类面前之前，我们能多大幅度地把这座数据山压缩下去。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">为衡量这一点，我们会持续追踪有多少 raw findings 能够通过各个验证阶段。得益于 Recon 阶段更好的上下文注入，我们的初始验证拒绝率从 40% 降到了 11%；与此同时，高完整度 findings 的占比从 35% 提升到 58%（对应约 12,057 条历史 findings）。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">下面是截至这篇文章写作时，从原始候选项到可执行 findings 的全生命周期拆解。</span></p><figure style="width: 728px;margin: 0px auto;"><p><picture style=";"><span leaf=""><img data-aistatus="1" alt="BLOG-3344 2" class="rich_pages wxw-img" data-ratio="0.7148148148148148" data-type="jpeg" data-w="1080" height="1041" sizes="100vw" style="display: block;max-width: 100%;height: auto;margin: 0px auto;" title="BLOG-3344 2" width="1456" data-imgfileid="100001418" src="https://wechat2rss.xlab.app/img-proxy/?k=30123494&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fsz_mmbiz_jpg%2FNBNzfW7CHBUuSt6vJARn472uja9Q1cTFicNyWY2zicfTNdQMsPPvdibiaTDg1vzpmtWH2aFZaMah3mnRpQzhDXtKECicbLAOqJ0wLeFkkwdzKicmg%2F640%3Fwx_fmt%3Djpeg%26from%3Dappmsg"/></span></picture></p><p><picture style=";"><span leaf="">Vulnerability Discovery Harness（VDH）：</span></picture></p></figure><ul style="margin-top: 0px;padding: 0px;" class="list-paddingleft-1"><li style="margin: 8px 0px 0px 32px;"><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);box-sizing: border-box;padding-left: 4px;"><span leaf="">Raw candidates：独立验证之前，discovery harness 输出的全部内容。</span></p></li><li style="margin: 8px 0px 0px 32px;"><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);box-sizing: border-box;padding-left: 4px;"><span leaf="">Needs repro：看起来可信，但需要人工复现后才能确认的 findings。</span></p></li><li style="margin: 8px 0px 0px 32px;"><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);box-sizing: border-box;padding-left: 4px;"><span leaf="">Rejected at validation：validator 推翻了威胁模型、exploit path、受影响代码或证据。</span></p></li><li style="margin: 8px 0px 0px 32px;"><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);box-sizing: border-box;padding-left: 4px;"><span leaf="">Duplicates：被折叠到同一 harness 生成的另一条 finding 上的候选项。</span></p></li><li style="margin: 8px 0px 0px 32px;"><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);box-sizing: border-box;padding-left: 4px;"><span leaf="">Survived validation：通过独立验证 gate，并进入 VVS 的 findings。</span></p></li><li style="margin: 8px 0px 0px 32px;"><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);box-sizing: border-box;padding-left: 4px;"><span leaf="">Bugs that went elsewhere：被有意路由到这条流程之外的 findings。</span></p></li></ul><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">Vulnerability Validation System（VVS）：</span></p><ul style="margin-top: 0px;padding: 0px;" class="list-paddingleft-1"><li style="margin: 8px 0px 0px 32px;"><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);box-sizing: border-box;padding-left: 4px;"><span leaf="">Another vulnerability harness：输入同一个验证系统的其他自动化来源。</span></p></li><li style="margin: 8px 0px 0px 32px;"><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);box-sizing: border-box;padding-left: 4px;"><span leaf="">Total bugs in system：ingest 之后形成的合并池。</span></p></li><li style="margin: 8px 0px 0px 32px;"><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);box-sizing: border-box;padding-left: 4px;"><span leaf="">Duplicates：dedup 阶段识别出的、已由另一条 canonical finding 或 ticket 覆盖的 findings。</span></p></li><li style="margin: 8px 0px 0px 32px;"><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);box-sizing: border-box;padding-left: 4px;"><span leaf="">Wrong repo / other / not a risk：噪声桶，包括归属仓库错误、纵深防御问题，或潜在但暂不构成风险的问题。</span></p></li><li style="margin: 8px 0px 0px 32px;"><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);box-sizing: border-box;padding-left: 4px;"><span leaf="">Bugs sent to teams：整理完成、可以交给工程团队修复的干净 findings。</span></p></li><li style="margin: 8px 0px 0px 32px;"><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);box-sizing: border-box;padding-left: 4px;"><span leaf="">Judged Internet-exploitable：现实攻击者可以在生产环境触发的高优先级 findings。</span></p></li><li style="margin: 8px 0px 0px 32px;"><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);box-sizing: border-box;padding-left: 4px;"><span leaf="">Not judged Internet-exploitable：优先级较低但仍可执行的 bugs，包括生产问题、依赖风险或配置错误。</span></p></li><li style="margin: 8px 0px 0px 32px;"><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);box-sizing: border-box;padding-left: 4px;"><span leaf="">Final severity split：用于给工程团队分配优先级的最终严重性分类。</span></p></li></ul><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">harness 的核心指标并非某个猜测出来的召回率，而是尽可能让真实人类面前的未确认 findings 数量接近于零。整个架构需要成为一个持续压缩噪声的过滤漏斗。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">在 VDH 生成的 20,799 条 raw candidates 中，最终只有大约 12,057 条通过了验证。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这些 findings 被推入 VVS，并与来自另一个 harness 的 findings 汇合后，中央池达到 13,841 条。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">Dedup agent 把其中 5,442 条作为重复项折叠掉。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">另外 1,154 条被路由到“wrong-repo”或“low-risk”队列，并在适当情况下重新回收到系统中。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">最终，留下 7,245 条可执行 findings，交给工程团队处理。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">传统合规规则通常完全基于静态 CVSS 分数，规定任意化的修复窗口，例如“所有 High 必须在 30 天内修复”。我们的 contextual judgment 层会把这个合规勾选项，转化成真正的风险管理。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这套架构能够把 findings 追溯到源头，这意味着修复一个根因，可以解决一整组 findings，而不是逐个打补丁。VDH 的系统性能也会通过另一种方式衡量：把代码仓库拆成（area × attack-class）单元，然后迭代运行 Gapfill agent，直到它不再产生 findings。每当我们更新底层 prompt 时，都会用一个留出的 repository 进行测试，看总覆盖单元数量是否真的发生变化。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">harness 还接入了自动化健康信号，用来在流水线早期发现系统故障。如果一次 hunt 结束得异常快，而且没有派生 sub-hunts 或 gap tasks，通常说明某个依赖崩溃了，而不是代码库真的干净。为解决这个问题，系统会把任何以零 findings 结束的 Hunter agent 标记为 “shallow”，并立即重新排队运行。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">最后，系统的稳健性还来自前面提到的独立分诊阶段。通过使用不同模型和不同逻辑权重，对所有提交进行重新判断，我们可以确保发现过程之外存在一层不带偏见、对抗性的验证。这一信任层与具体用于 discovery 的模型解耦，因此无论当前使用的是哪一个模型，它都能持续存在。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">我们的北极星指标：衡量真实世界的推进速度</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">每个代码库都有些不同。为了说明这套系统在真实世界里到底如何运行，我们基于一次标准仓库运行，整理了一组更接近现实的基准。需要注意的是，这代表的是单个代码仓库的一次扫描；随着持续运行的整组代码仓库循环不断去重、过滤和回收 findings，长期候选项总量大约会减少 65%。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">通过自动化打补丁节省工程时间：相比静态基线，我们更关注流水线的技术吞吐量、处理速度，以及它消除人工分诊瓶颈的能力。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">Initial Validation Cut：对于一个标准代码仓库（约 3 万行代码），系统会产生 100 条初始 findings；一次完整运行需要 3 到 4 小时，并且全程保持高度聚焦的上下文窗口。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">Compression：Deduplication 和 Contextual Judgment 层会并行处理这些候选项。3 小时内，系统会把这批 findings 从约 100 条 raw candidates 压缩和精炼为 80 个彼此不同的高可信漏洞。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">Remediation：自动化 Fixer 会以平均每个漏洞 5 分钟的速度处理这 80 个不同漏洞。总体来看，系统可以在大约 14 小时内完成发现、验证、去重，并打开可工作的 pull requests。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">缩短关键缺陷的 mean-time-to-resolve：当然，你不能一次性把 80 个 patches 全部倒进生产环境，否则很容易把系统弄坏。为了保证部署安全，我们的系统采用分层 rollout。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">Critical Exposure Containment：系统会先隔离 critical、high 和 exploitable 漏洞，平均大约是 80 个中的 10 个。我们会让这些问题进入快速人工审查，并纳入发布周期，使它们在 5 天内完成生产环境修复。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">Incremental Hardening：剩下的潜在风险、轻微配置异常和较低优先级漏洞，会在 15 到 20 天窗口内逐步进入生产环境，以保证平台稳定性。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">我们如何处理这些补丁</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这些 findings 来自一个隔离、受限的研究实验，目的是对我们的代码进行压力测试。它们并不代表当前生产环境中仍然存在活跃、未修复的漏洞。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">由于 harness 会在测试环境中持续运行，等你读到这里时，这些具体数字已经完全过期。流水线发现的每一个漏洞，都附带了一个可运行测试用例，用来演示问题，同时也带有一个草拟 patch。我们的安全团队正在系统化处理这些报告，并应用必要修复。这意味着你每天使用的 Cloudflare 产品，已经在主动针对这些攻击向量进行加固。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span style=";"><span leaf="">随这篇文章一起，我们也发布了最初用于开发 harness 的 skill。发布前我们对它做了一点清理，让它更容易理解和集成，但 skill 本身基本保持不变。希望 harness 本身也会很快发布。它可以成为你构建自己的 vulnerability harness、自己的 skill，或任何最适合你需求的系统的起点：</span></span><span leaf="">github.com/cloudflare/security-audit-skill</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span style=";"><span leaf="">如果你的团队也在解决类似问题，并希望交流经验，可以通过 </span></span><span leaf="">security-ai-research@cloudflare.com</span><span style=";"><span leaf=""> 联系我们。</span></span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">所有这些都还没有完成。我们一直在改系统，它离一门完美科学还很远。但现在，原始候选 findings 已经变得便宜，真正值得做的工作，是把它们转化成可靠、可验证的代码修复。</span></p><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);"><span leaf="">构建自己的 harness，意味着接受 AI 模型本身是易变的，但你的编排层不必如此。通过把安全逻辑从任何单一供应商中解耦出来，强制引入对抗性验证，并自动化你的分诊流水线，你就可以把一座由 LLM 噪声构成的大山，转化为一个可靠的、覆盖整组代码仓库的防御引擎。</span></p><p style="display: none;"><mp-style-type data-value="3"></mp-style-type></p>



<p><a href="https://wechat2rss.xlab.app/link-proxy/?k=b820938e&amp;r=1&amp;u=https%3A%2F%2Fmp.weixin.qq.com%2Fs%3F__biz%3DMzUzOTI4NDQ3NA%3D%3D%26mid%3D2247485070%26idx%3D1%26sn%3Dd4483a9d95a91235a17c255b383cdcbc">跳转微信打开</a></p>
]]></content:encoded>
      <pubDate>Sat, 27 Jun 2026 00:05:00 +0800</pubDate>
    </item>
    <item>
      <title>[译苑雅集vol. 11]程序员开始焦虑失业，哲学家却被 AI 公司抢人</title>
      <link>https://mp.weixin.qq.com/s?__biz=MzUzOTI4NDQ3NA==&amp;mid=2247485061&amp;idx=1&amp;sn=e3957fe687948d4d0d64501ee81a647a</link>
      <description>大型 AI 实验室正在招聘哲学家。哲学正在进入模型推理、安全宪法、价值取舍、自动驾驶和武器系统，成为训练与部署 AI 的实用工具。</description>
      <content:encoded><![CDATA[<p><span>四楼南侧东</span> <span>2026-06-25 23:05</span> <span style="display: inline-block;">北京</span></p>






  
  <p><img src="https://wechat2rss.xlab.app/img-proxy/?k=b604530c&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fmmbiz_jpg%2FNBNzfW7CHBWc16TlfmLsibcxXRXtWhgGOZA0ZUz4aaaM0D9hoHvTvHNicbPuLkZIYJgFdnV0hMQnHkx54iaibMQsJQtfFGjpXBZn2WvPkibmrZ6I%2F0%3Fwx_fmt%3Djpeg"/></p>
  <p>大型 AI 实验室正在招聘哲学家。哲学正在进入模型推理、安全宪法、价值取舍、自动驾驶和武器系统，成为训练与部署 AI 的实用工具。</p>
  <p style="text-align: center;" nodeleaf=""><img data-aistatus="1" class="rich_pages wxw-img" data-ratio="0.562962962962963" data-s="300,640" data-type="png" data-w="1080" type="block" data-imgfileid="100001409" src="https://wechat2rss.xlab.app/img-proxy/?k=02e8b923&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fmmbiz_png%2FNBNzfW7CHBWZib2TGPbzPW07NIwO3QiaN8DVotNLrSRhDlCGsDYcs6Lqw2pD0VIticEZf2D2SjlsmWTjNYw1ZN3THbzfjib6ykjGfbYZicVfZr3I%2F640%3Fwx_fmt%3Dpng%26from%3Dappmsg"/></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);" data-pm-slice="0 0 []"><span leaf="">作者：The Economist</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">时间：2026年06月24日</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">原文：<a href="https://www.economist.com/science-and-technology/2026/06/24/why-big-ai-labs-are-hiring-so-many-philosophers" target="_blank">https://www.economist.com/science-and-technology/2026/06/24/why-big-ai-labs-are-hiring-so-many-philosophers</a></span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">十年前，随着 AI 革命开始加速，艺术和人文学科的学生被告知，如果想让自己更有就业竞争力，就应该“学会编程”。这条建议现在看来可能并不高明。如今，对 AI 可能抢走饭碗感到紧张的，反而是程序员。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">他们或许可以考虑去学哲学。纽约联邦储备银行今年早些时候发布的数据显示，在美国，哲学专业毕业生比计算机科学专业毕业生更容易找到工作。2024 年，也就是目前有数据可查的最近一年，计算机科学专业毕业生失业率为 7%，哲学专业毕业生则为 5.1%。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">其中不少人就是被 AI 公司抢走。耶鲁大学哲学家 Luciano Floridi 说，有些学生还没毕业就已经拿到工作机会。学者也同样在流动，Floridi 形容，哲学系人才流失的规模已经像是一次“失血”。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">哲学能教给 AI 研究者的一些东西，其实相当古老。“苏格拉底方法”会通过佯装无知和连续追问，澄清含义、发现矛盾，并揭示一个观点可能带来的后续影响。今天的许多 AI 系统都有迎合用户的倾向，慕尼黑大学研究哲学与 AI 的专家 Jörg Noller 说，经过苏格拉底方法训练的模型，不那么热衷于取悦人类，更愿意追求真相。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">还有“苏格拉底式无知”这个观念，在《申辩篇》中，柏拉图笔下的苏格拉底声称，他的智慧主要在于意识到自己有多么无知。把这种谦逊植入模型，可以帮助限制过度自信。过度自信是一个常见缺陷，Noller 称之为“AI 不成熟”。伦敦 AI 实验室 Google DeepMind 的高级哲学家 Iason Gabriel 认为，整个行业的幻觉现象有所下降，可以归因于这类努力。更广泛地说，他认为，哲学训练是改善 AI 长推理过程，也就是所谓“思维链”的“一种强有力机制”。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">哲学训练还会以更具体的方式影响模型的立场。特拉华大学技术哲学家 Thomas Powers 说，如果把 约翰·洛克 的著作输入一个 AI 法律助手，它就会倾向于把强有力的财产权视为政治自由的基础。如果你不喜欢这些原则，模型制造商那里也有别的选择。美国计算巨头 IBM 的 Granite 系列模型带有一些调节旋钮，可以让企业客户更好地把输出结果与自己的公司哲学对齐。IBM “负责任 AI” 的负责人 Francesca Rossi 说，这些功能可以让用户在不同哲学取舍之间选择平衡点，比如个人自由意志与社会稳定和谐。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">哲学也能帮助 AI 安全。研究人员已经记录下 AI 模型中的各种令人不安的行为，包括试图逃避监督，甚至勒索用户。模型制造商试图抑制这类不当行为，其中一种方法叫作 AI 宪法主义。它的做法是，从具有法律或道德权威的哲学文本中提取规则和原则，并以此给模型搭建一套约束框架。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">总部位于旧金山的 AI 实验室 Anthropic 就是这一方法的支持者之一。它为 Claude 模型制定的宪法，曾纳入来源相当多样的材料，包括康德的著作、Apple 的服务条款和《世界人权宣言》。最新一版由 Anthropic 首席哲学家 Amanda Askell 主导，于 1 月 21 日发布。Anthropic 内部一些员工给这份 78 页的宪法起了个昵称，叫 Claude 的“灵魂文档”。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">不过，最大的问题在于，这些宪法一开始究竟应该放进什么样的规则。哲学家们把注意力集中在两种主要伦理框架上。第一种是义务论，康德是这一传统的重要代表。义务论会施加严格规则，禁止说谎、胁迫，以及把人只当作手段而不是目的来对待，即便这样做是为了更大的善。Anthropic 的宪法纳入了许多义务论约束，前述特拉华大学技术哲学家 Powers 博士认为，这些约束可以让 AI 行为更加一致。对于要部署到家庭和公共空间里的机器人来说，这是一项优点。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">以义务论看待世界的模型还有其他好处。其中之一是更诚实，而诚实正是 Claude 被广泛注意到的一个特征。牛津大学哲学家 Nick Bostrom 说，更诚实的模型更不容易误导用户。另一家硅谷实验室 Inflection AI，则把义务论约束施加到它的 Pi 聊天机器人上。Pi 被设计用来提供情感支持。公司负责人 Sean White 说，Pi 很擅长识别那些可能伤害自己或他人的用户。Floridi 说，义务论宪法还有助于满足法律合规要求。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">AI 哲学家感兴趣的另一种伦理路径叫后果主义。它通过权衡成本与收益来决定应该怎么做。更接近后果主义的模型包括 OpenAI 的 ChatGPT 和 Google 的 Gemini。Google 的 AI 模型被设计成要产生“可能的总体收益，并且这些收益大幅超过可预见风险”，这就是一种典型的后果主义目标。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">后果主义算法在自动驾驶汽车软件中也至关重要：如果事故无法避免，系统就必须决定哪一种撞车方式造成的损害最小。自动驾驶公司 Waymo 的高级工程师 Chris Gerdes 说，趋势是让驾驶软件变得更加后果主义。后果主义也是 AI 武器系统 的核心。美国军方联合人工智能中心前负责人 Jack Shanahan 说，军事目标必须与可能造成的平民死亡相权衡。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">棘手问题比比皆是，而这正是哲学家喜欢的问题类型。是否存在一些情况下，义务论规则应该被推翻？当后果并不清楚时，应该怎样做决策？AI 系统是否应该考虑动物福利，或者环境状态？Nauto 是一家为卡车和其他商用车开发 AI 安全系统的公司，它的负责人、哲学家 Stefan Heck 问道：优先保护年轻行人，而非年长行人，在道德上是否可以接受？他预计，伦理上高度敏感的诉讼会出现。毕竟，后果主义算法明确允许一种伤害发生，只要它的设计目标是避免更糟糕的伤害。</span></p><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);"><span leaf="">批评者担心“道德去技能化”：如果计算机越来越多地做出伦理判断，人们会不会变得不那么愿意自己做判断？路易斯维尔大学 AI 理论家 Roman Yampolskiy 认为，道德“在历史上并不稳定，在文化上各不相同，在策略上可以被操纵，而且往往只有事后才看得清楚”。</span></p><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);"><span leaf="">失业的程序员们可以留意一下：AI 哲学家看起来并不缺工作。</span></p><p style="display: none;"><mp-style-type data-value="3"></mp-style-type></p>



<p><a href="https://wechat2rss.xlab.app/link-proxy/?k=cc558ca2&amp;r=1&amp;u=https%3A%2F%2Fmp.weixin.qq.com%2Fs%3F__biz%3DMzUzOTI4NDQ3NA%3D%3D%26mid%3D2247485061%26idx%3D1%26sn%3De3957fe687948d4d0d64501ee81a647a">跳转微信打开</a></p>
]]></content:encoded>
      <pubDate>Thu, 25 Jun 2026 23:05:00 +0800</pubDate>
    </item>
    <item>
      <title>[译苑雅集Vol. 10]别再比谁烧 token 多了：企业 AI 要算 ROT</title>
      <link>https://mp.weixin.qq.com/s?__biz=MzUzOTI4NDQ3NA==&amp;mid=2247485056&amp;idx=1&amp;sn=faa23efe0ce9856d7ee6b5618803ebdb</link>
      <description>本文提出 Return on Tokens（ROT）这个视角：企业 AI 的关键，是把 token 转化成低成本、可复用、可演化的业务软件，而非最大化消耗。</description>
      <content:encoded><![CDATA[<p><span>四楼南侧东</span> <span>2026-06-23 17:47</span> <span style="display: inline-block;">北京</span></p>






  
  <p><img src="https://wechat2rss.xlab.app/img-proxy/?k=ca5b48fe&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fmmbiz_jpg%2FNBNzfW7CHBUP7NXiaI7gURqzNBwRER8jOJyWLb6UdTibzZwcndmib8mnvskwpNP7kibNsADqT7iaAs79g49ZwNCSLRd46P1vJZLulPZE7YxJ3ZjE%2F0%3Fwx_fmt%3Djpeg"/></p>
  <p>本文提出 Return on Tokens（ROT）这个视角：企业 AI 的关键，是把 token 转化成低成本、可复用、可演化的业务软件，而非最大化消耗。</p>
  <p style="text-align: center;" nodeleaf=""><img data-aistatus="1" class="rich_pages wxw-img" data-ratio="0.562962962962963" data-s="300,640" data-type="png" data-w="1080" type="block" data-imgfileid="100001401" src="https://wechat2rss.xlab.app/img-proxy/?k=6f934bf5&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fsz_mmbiz_png%2FNBNzfW7CHBW0ITiaYGDMuaiaEZGhUAIhqZk7BAmRyHyYJibfCdf2sSHAv0ibFskRLIZWZGhKgM1yeuuFtT3mcNdcL9k1dE8tVO2FGgiaTDVU0NRg%2F640%3Fwx_fmt%3Dpng%26from%3Dappmsg"/></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);" data-pm-slice="0 0 []"><span leaf="">作者：Packy McCormick、Markie Wagner</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">时间：2026 年 06 月 11 日</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">原文：</span><span leaf=""><a href="https://www.notboring.co/p/return-on-tokens-rot" target="_blank">https://www.notboring.co/p/return-on-tokens-rot</a></span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);" data-pm-slice="0 0 []"><span leaf="">AI 的承诺在于，它会把业务变成软件，让业务能够通过数百万次微小迭代不断演化。那些美好、理想而复杂的东西，只能在漫长时间里的大量试错之后出现。完美无法被凭空建造，只能在不断试错中被发现。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">资本主义，本质上是一场组织层面的进化。数以百万计的企业在市场中竞争，拿出它们认为客户会需要的产品和服务。有些企业繁荣壮大，有些企业走向死亡。每家公司内部也在持续演化：人员来来去去，一次实验变成流程，流程沉淀成一张隐性知识之网；新产品被推出，旧产品被淘汰。正是这种持续演化，让我们拥有了今天的生活水平；未来世代回看今天，也会觉得我们仍然相当原始。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我离开研究世界，是因为我觉得那不是最值得攀登的山头。于是我走进美国各地，直接去做具体的工作。只有先弄清楚工作本身，我才能理解怎样让计算机去做这些工作，进而让人类能够指挥计算机，推动工作不断演化。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">在真正走出去之前，我已经有所怀疑；一路看下来之后，我更加确信：到目前为止，市场实现 AI 的方式是错的。它还不是终局。这套方式太浪费、太健忘，也太不精确。为了换来这点认识，我简直像是在撒哈拉沙漠里和恶魔搏斗了一场。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">Tokenmaxxing 显然走不远</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">所谓 tokenmaxxing，字面意思就是尽可能最大化个人或组织消耗的 token 数量。企业用排行榜追踪 token 消耗，再用小奖品奖励使用量最高的人。这是一场集体幻觉，像是一种商业化的 AI 精神错乱。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">tokenmaxxing 是一个实验室培养出来的超级 meme，而且传播效果可能比这些 AI 实验室原本期待的还要好。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">可以这样想象：Anthropic 和 OpenAI 分别推出了 Claude Code / Cowork 和 Codex 这类 Agent 产品。这些 Agent 基本上就是 AI 模型公司派进客户公司里的“员工”，还拿到了一张没有消费上限的公司信用卡。按照 tokenmaxxing 的逻辑，它们花得越多，就说明干得越好；而它们真正花钱的对象，正是它们的实际雇主，也就是 AI 模型公司。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">比如，Anthropic 把一批 Agent 送进 KPMG。KPMG 承诺一定的使用量，以换取折扣，也就是 token commits。KPMG 的员工被鼓励用 Agent 做任何他们能想到的事情，各种仪表盘也随之出现。于是，这些可以被理解成“拿着 KPMG 无限额信用卡的 Anthropic 数字员工”的 Agent，开始尽情消耗 token，把账单刷到痛快为止。那些指挥 Agent 使用最多 token 的员工，则被表彰为 AI Innovators。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">当然，也有人意识到这是一场幻觉。他们会问：“可是这些 Agent 到底有没有做出什么有用的东西？它们是不是只是在做仪表盘？能不能请谁给我看一个真正有用的 Agent 产出？”但这些少数清醒的人，很快会遇到一句杀伤力极强的反驳：“Skill Issue。”</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">他们被告知，有些人正在用 Agent 构建极其有价值的东西。这有点像有人说自己在夏令营有一个超级火辣的女朋友，只是你从来没见过。如果你自己做不到同样的事，那好吧，欢迎来到 Permanent Underclass。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">有一段时间，所有人都信了。市场激励公司消耗 token，于是董事会激励管理层消耗 token，管理层激励经理消耗 token，经理又激励员工消耗 token。没有人有动力站出来说，这些 token 并没有在做什么有用的事情。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我总是在和这些人聊天，而每家公司都会出现某个版本的同一段对话。负责 AI 团队的人会说：“这个季度我们取得了巨大进展。我们花了 5000 万美元买 token。”然后所有人点头、鼓掌。“使用量上来了。我们做了 3000 个 Agent。我们交付了 1000 万行代码。”你听完只想问：什么？然后我会问：“那个反欺诈 Agent，你们测过准确率吗？”对方说：“测了……大概 50%。”可人能够做到 99% ！</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">但是模型还在进步，所有人都在 tokenmaxxing，而你也不想显得像一个抗拒技术的人。于是，你继续懒散地把 Agent 扔向所有问题，然后指望它们自己学会。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">顺便说一句，这一切发生的时候，正好也是 AI 模型公司从订阅制收入转向按量消耗收入的时候，企业几乎没有时间准备。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">所以，token 使用量，以及由此带来的 AI 模型公司收入，出现抛物线式增长，也就一点都不奇怪了。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">最终打破这场幻觉的，是 Uber。上一轮周期里，Uber 曾是 VC 补贴需求的标志性公司。它的 CTO 说，公司到 4 月就已经烧完了 2026 年 Claude Code 的 token 预算。5 月，Uber COO 又说，公司越来越难证明 AI 支出的合理性，因为 AI 消耗和实际交付功能之间的联系“还不存在”。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">接下来发生的事情，有点像电影《贱女孩》里的一个场景：Tina Fey 问学生们，如果她们觉得自己曾被 Regina George 伤害过，就请举手。先是一只手举了起来，然后所有人的手都举了起来。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span style=";"><span leaf="">有位顾问说，他的客户不小心在 Claude Code 上烧掉了 5 亿美元。</span></span><span leaf="">Amazon 关闭了自己的 AI 排行榜</span><span style=";"><span leaf="">。Legora CTO Jacob Lauritzen </span></span><span leaf="">告诉 Harry Stebbings</span><span style=";"><span leaf="">，token 排行榜“会导致 tokenmaxxing，也就是人们为了显得自己做得好，就去白白烧 token。这是一种非常愚蠢的做事方式。”Ramp 的 Veeral Patel 把它称为 </span></span><span leaf="">Token Casino</span><span style=";"><span leaf="">：“有用的软件，被包裹在一种让花钱看起来像进步的机制里。它从最古老的把戏开始：把钱抽象掉。”Palantir CEO </span></span><span leaf="">Alex Karp 对 TBPN boys 说</span><span style=";"><span leaf="">，tokenmaxxing 就像“一种色情成瘾”。</span></span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">甚至连 Sam Altman 这个重要的 token 供应商本人，也在 CNBC 上承认：“你会听到一些公司说，‘我在 AI 上花了很多钱，我知道一些很棒的事情正在发生，但我也知道这里面有大量浪费。而且，你知道，什么时候……我要等多久，才能真正看到它体现在收入里？我要等多久，才能真正把成本控制住？’” 他承认，这已经变成了一个“巨大问题”。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">问题在于，企业一直把注意力放在最大化 token 上，并假设 token = 价值。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">每个周期都有自己的愚蠢指标。19 世纪中期，市场把铁路里程当成未来垄断及其收益的代理指标，于是铁路公司竞相铺轨，很多线路甚至和竞争对手走在同一条路线上。21 世纪初，市场想要的是 eyeballs，也就是眼球，于是互联网公司拼命吸引眼球，再把它们端上广告主的餐盘。2010 年代，市场想要的是收入总额，于是 WeWork 这样的公司交出了好看的收入总额。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这一轮周期的愚蠢指标，就是 tokenmaxxing。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这并不意味着 token 没有价值。Cornelius Vanderbilt 的 New York Central 最终变得非常有价值，Pennsylvania Railroad 也是如此。Google 和 Facebook 把眼球转化为现金流的能力，可能比历史上任何人把任何东西转化为现金流的能力都强。Uber 最后也把收入增长转化成了市场主导地位，并在 2025 年形成了 100 亿美元自由现金流。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">问题始终在于：这个东西能不能产生回报？</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">对于 token 来说，问题就是：你的 Return on Tokens，也就是 ROT，到底是多少？</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">Return on Tokens</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">当你投资一台新机器时，你会期待它产生回报。当你招聘一名员工时，你也会期待他产生回报。商业的过程，就是不断做出大大小小的投资，并期待这些投资创造出超过成本的价值。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">token 也应该接受同样的标准。</span></p><pre style="background: none 0% 0% / auto repeat scroll padding-box border-box rgb(255, 255, 255);padding: 20px;margin-top: 0px;margin-bottom: 20px;outline: rgb(230, 230, 230) solid 1px;outline-offset: -1px;border-radius: 12px;box-sizing: border-box;"><code style="white-space: pre-wrap;font-family: monospace;font-size: 13px;line-height: 24px;font-weight: 500;color: rgb(54, 55, 55);"><code style="white-space: pre-wrap;font-family: monospace;font-size: 13px;line-height: 24px;font-weight: 500;color: rgb(54, 55, 55);"><span leaf="">Return on Tokens = （产出价值 - token 成本）/ token 成本 x 100</span></code><br/></code><br/></pre><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">由此看，提高 ROT 有两条路径：你可以用 token 创造更有价值的东西，也可以减少 token 消耗。理想状态下，是用更少的 token 创造更多价值。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">企业现在首先关注的，是减少消耗。原因很简单：相比衡量产出价值，衡量支出要容易得多。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">当 tokenmaxxing 的幻觉被打破后，更冷静的人开始讨论 “routing”，也就是模型路由，把它作为降低成本的办法。需要高强度推理的事情，用 Anthropic 和 OpenAI 最好的模型；大部分普通工作，则交给便宜的中国开源模型。Coinbase CEO Brian Armstrong 最近的公开说法，就是这种逻辑的典型例子。</span></p><figure style="width: 728px;margin: 0px auto;"><p><picture style=";"><span leaf=""><img data-aistatus="1" class="rich_pages wxw-img" data-ratio="0.9129955947136564" data-type="png" data-w="908" height="829" sizes="100vw" style="display: block;max-width: 100%;height: auto;margin: 0px auto;" width="908" data-imgfileid="100001403" src="https://wechat2rss.xlab.app/img-proxy/?k=930bdf8d&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fmmbiz_png%2FNBNzfW7CHBXN6KDM9HrhSHOEd8vvpdF44gGiafVShHyJwHuMAwTicGNOz6BzBYg4KupKnKXLKBkEibp3knMsSn3ddpoYCwEGBtboibJwbap3gwA%2F640%3Fwx_fmt%3Dpng%26from%3Dappmsg"/></span></picture></p></figure><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span style=";"><span leaf="">你也可以在 </span></span><span leaf="">OpenRouter AI Model Rankings</span><span style=";"><span leaf=""> 里看到这种变化。转向中国模型的趋势，实际上和按消耗计费的普及几乎同步出现。当然，这里的用户本来就是一个自我筛选过的群体，他们已经在思考怎样把不同任务路由给合适的模型。其他人现在也在匆忙做同样的事。</span></span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这是一个不错的开始。但让 Agent 消耗 token，不管是美国模型的 token 还是中国模型的 token，从零开始推导所有事情，也仍然到不了终局。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">因为你知道什么比中国模型还便宜吗？代码。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">代码，那种古老而朴素的确定性代码，不仅便宜，而且更适合大多数具有经济价值的工作。这个教训，我们其实早就学过。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">过去，公司雇佣人类来完成各种重复性任务。在 “computer” 变成数字机器之前，computer 指的就是做人肉计算的人，也就是计算员。</span></p><figure style="width: 728px;margin: 0px auto;"><p><picture style=";"><span leaf=""><img data-aistatus="1" class="rich_pages wxw-img" data-ratio="0.5814977973568282" data-type="png" data-w="908" height="528" sizes="100vw" style="display: block;max-width: 100%;height: auto;margin: 0px auto;" width="908" data-imgfileid="100001405" src="https://wechat2rss.xlab.app/img-proxy/?k=e0aa49a1&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fmmbiz_png%2FNBNzfW7CHBV8JrYBUbBbJicviaYuFlKMsEVy5JjM7tNbvq67NCS8lnccxpvyY8sMVboiaVtxkEQq2ErNQ5ib8GeCgmCxl0zSApg4DibqNhwEYcLM%2F640%3Fwx_fmt%3Dpng%26from%3Dappmsg"/></span></picture></p></figure><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><em style=";"><span style=";"><span leaf="">When Computers Were Human, </span></span><span leaf="">NASA</span></em></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">大约半个世纪前，我们开始把人类做的重复性任务交给软件，比如计算导弹轨迹，或者计算一家公司的利润。代码比最可靠的人类计算员还要可靠。它不会犯错，能够立刻给出答案。无论在世界上的哪个地方、哪个时间，只要你在 Excel 表格中把同样的数字和同样的公式输入到同样的单元格里，它就会吐出同样的数字。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">然后，我们有了 Agent，却忘掉了这个教训。我们开始觉得，既然所有人都在这么做，那就应该把这些“类人”的东西扔向所有问题。Agent 在某些事情上确实很强，但在很多其他事情上，它们并不适配。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">企业没有从 token 上获得正向 ROT，并不奇怪。所有能做的仪表盘都已经被做成了仪表盘，现在它们又开始派 Agent 去做本来属于软件的工作。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">当然，也可以提出一种说法：很多公司之所以拿不到正向 Return on Tokens，是因为它们还不知道怎样使用 token，或者还没有把公司重新架构成 AI 原生的形态。或许也正因为如此，Anthropic 和 OpenAI 都推出了咨询子公司，帮助企业更好地部署 token。确实也有一些在 AI 时代诞生的创业公司，看起来非常有效地使用了 token，这一点也体现在它们超音速般的收入增长上。如果旧经济企业做不出 ROT，那好吧，这就是破坏性创新。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">虽然不同公司部署 AI 的能力确实有差异，但我们认为，负 ROT 背后存在更根本的结构性原因：对大多数工作来说，Agent 是错误的架构。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">Agent 也走不通，至少对大多数工作来说</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">Agent 的负 ROT，有三个结构性原因。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">第一，Agentic 架构无法以真实经济工作所要求的“几个 9”的质量水平，完成长时间运行的工作。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">Agent 依赖即兴发挥。它们被一次次重新派到重复性任务上，好像每天都是上班第一天，这会伤害持续稳定的准确性。对于新功能、原型或仪表盘，80% 的准确率也许可以接受。但对于支撑经济运转的真实重复性工作，比如欺诈检测或承保决策，80% 的准确率等于 0% 可用。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">第二，工程师不知道该构建什么，因为他们并不亲自做那些工作。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我们正在讨论的大多数流程型工作，通常由两部分组成：一部分是书面规则，Agent 可以读取；另一部分则是大约 3000 条存在于人脑中的隐性规则和子规则。这些知识分散在全国各地的办公室里，远离工程师在旧金山的办公桌。AI 只能演化它能够触及的东西，这也是为什么它在写代码上表现很好，却在企业里很大程度上没能做出真正有用的事情。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">第三，原罪在于没有目标。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">如果人没有目标，Agent 也就没有目标，最后整个系统什么结果也达不成。没有一个清晰的目标，代码无论是人写的，还是 Agent 生成的，最终都会退化成一团垃圾，因为没有一种力量去评估什么是好、什么是坏。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">从偷懒的角度看，Agent 有一个很诱人的地方，尤其是在组织还鼓励你多花 token 的时候：你可以在自己并不确切知道要解决什么问题的情况下，直接把 Agent 放出去。它可以围绕一个模糊指令转上一阵，带回一个还不错但不完美的东西，然后再出去继续转。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这个过程会推动更多 token 消耗，却不创造任何价值。这是一条通向负 ROT 的快车道。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">人们正在为 Agent 寻找新的事情可做，背后的假设是：AI 对代码做过的事，也会对其他一切事情重演。但它未必需要这样。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">有相当多的工作，最适合的完成方式仍然是朴素的老式代码。过去的挑战在于，直到最近，我们还没有足够多的工程师，把每家企业所做的每一件事都变成代码，并随着情况变化不断更新。现在我们有了。AI 让写代码变得微不足道。所以，只要我们能把知识从人的脑子里取出来，就能把业务变成代码。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">AI 应该做编译器，而非运行时</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">基本上，软件的工作分成两步：思考和执行。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">第一步是思考。你拿到一段软件应该实现的目标和需求，然后把它编译成计算机可以运行的代码。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">第二步是执行，而且是一遍又一遍地执行。每当这段软件需要完成那件事时，代码就会以低成本、可预测，或者说确定性的方式运行。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">计算机科学对编译器有精确定义，但你也可以把一家软件公司，或者一名软件工程师，理解成一种编译器。他们接收目标和需求，把它们转化成代码。然后，客户购买他们构建出来的代码，一遍又一遍地运行。这就是零边际成本软件业务的美妙之处，也是为什么公司可以把耗费数百万美元开发出来的软件，以每月 20 美元的价格卖出去，同时仍然获得令人垂涎的利润率。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">今天，大多数人思考和使用 Agent 的方式，是让它们同时替代软件公司和软件本身。这种理解是错的。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">Agent 应该替代软件公司。它们应该接收用英语，或者任何其他语言表达的目标和需求，然后把它们转化成可以一遍又一遍确定性运行的代码。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">思考很贵，但很少发生。执行很便宜，而且会一直发生。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">Agent 应该负责思考，代码应该负责执行。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">对于大多数经济工作，你想要的模式应该是：用人类弄清楚规则，用 AI 把规则变成代码，然后让这些代码以接近零 token 成本的方式一直运行。只有当规则发生变化时，才把 AI 再请回来。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">为什么要用 prompt 去做两个数字相加？写一行 Python 就行了，兄弟。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">当前 AI 实施中的 Thinking-Doing Ratio，也就是思考-执行比，大约是 1000:1。这并不奇怪。旧金山是一个“思考”之城。Anthropic 的帽子上就只写着一个词：“thinking”。</span></p><figure style="width: 728px;margin: 0px auto;"><p><picture style=";"><span leaf=""><img data-aistatus="1" class="rich_pages wxw-img" data-ratio="0.7528248587570622" data-type="png" data-w="708" height="533" sizes="100vw" style="display: block;max-width: 100%;height: auto;margin: 0px auto;" width="708" data-imgfileid="100001404" src="https://wechat2rss.xlab.app/img-proxy/?k=32216b4f&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fsz_mmbiz_png%2FNBNzfW7CHBW0k9AA1uSUGvnaDEhNgKHHQO9q82SicXo4gCNhXo18ZhFYy6W0w52t1oBz99cicW045XeNZnlibzHmNmsZ4WiaZLBBRBfuWDBRtGE%2F640%3Fwx_fmt%3Dpng%26from%3Dappmsg"/></span></picture></p></figure><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><em style=";"><span style=";"><span leaf="">Anthropic thinking cap + OpenAI Token Appreciation Award, </span></span><span leaf="">Business Insider</span></em></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">硅谷构建 AI 时，假设工作主要是思考。但工作主要是执行。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">聊天是少数例外，因为在聊天里，你真的不知道下一步会发生什么。所以，也许客户支持业务仍然会持续消耗 thinking tokens，尽管即使是客户支持 Agent，也会把复杂问题转给人类。除此之外，企业里几乎没有什么工作像是在持续即兴发挥。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">所以，我们现在用 Agent 来执行，但 Agent 是执行领域里的 BlackBerry。它可能是一代过渡性设备，却不会是五年后企业运行大多数工作的地方。那些工作会在 Agent 写出来的确定性代码里完成。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">Agent 的角色，是编译成代码，而非每天运行并执行工作。这也意味着，它更像资本支出（CapEx），而非运营支出（OpEx）。你以为未来会由 AI 运行的东西，实际上都会由代码运行。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">所有人都以为，世界即将发生的变化是 AI 会变成一个人。但真正的变化，是业务会变成一段软件。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span style=";"><span leaf="">这就是我们在 </span></span><span leaf="">Poetic</span><span style=";"><span leaf=""> 正在构建的世界。</span></span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><span leaf="">把业务变成会演化的软件</span></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我们正在构建 tokenmaxxing 的解药：一种能够自我 tokenminn 的软件。</span></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">Poetic 是一种新类别的软件：像 AI 一样自适应，像代码一样可靠。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我们把 AI 当作编译器。我们会先吸收一家企业已经写下来的所有流程，然后去内布拉斯加、普罗维登斯，或者任何真正发生工作的地方，坐在员工身边，一遍又一遍地问：“你刚才做了什么？”“你为什么这么做？”通过数百次这样的追问，学习每家公司赖以运转的数千条隐性规则。然后，我们把这些规则变成代码。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">代码就是运行时。当世界保持不变时，这些代码每次都执行完全相同的步骤。当世界发生变化时，它会学习、重新生成、围绕目标进行自我测试，然后运行新的代码，直到世界再次发生变化。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">结果是，token 使用量降低 100 倍，同时在复杂任务上达到多个 9 的准确率。换句话说，你花出去的每一个 token 都能多做 100 倍的事情，而且做对。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">产出的价值会上升，因为 Poetic 做的是你的企业确实需要反复完成的事情。token 成本会下降，因为 Poetic 只在世界发生变化时使用 token。两者结合起来，Poetic 能够产生清晰、可衡量的 Return on Tokens。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我们今天已经在为 AIG、SoFi 和 Chime 这样的公司做这件事。AIG CEO Peter Zaffino 说，Poetic 已经“在持续数小时的流程中实现了 99% 以上的质量结果，交付了真实的企业价值”。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这些公司的领导者相信我们所相信的事情：每一家企业都必须以软件企业的方式被重新创立。未来十年的故事，就是这些新企业的开始。有些会是真正从零开始的新企业；另一些则是已经存在数百年的企业，但它们有勇气重新发明自己。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">所有人都在谈论，工厂只有围绕电力重新配置之后，才能真正受益于电力。但他们接着给出的 AI 版本，往往变成了“所以，那些围绕大量消耗电力来解决问题的新企业会赢”。关键动作其实是把业务重构成代码。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">要做到这一点，需要脚踏实地的地面战。你必须深入企业的内部，无论它们在哪里，理解它们如何运转，并把它们的逻辑迁移进程序里。我们需要有人真的走到明尼苏达去，问一句：你们这些人一天到晚到底在干什么？</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我们团队的大多数人都是工程师，其中很多来自 Palantir。他们会一次在客户现场待上数周，向客户学习，钻进最细碎、最具体的细节里。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">“Social Engineer” 这个词名声不太好，尤其在安全语境里容易让人想到社会工程攻击。但相比那些整天坐在桌前给 Claude 写 prompt 的工程师，他们才是最社会化的工程师。理解人、理解商业、理解 AI 的工程师，将会统治世界。如果这听起来像你，欢迎加入 Poetic。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这是一份艰苦的工作。但 AI 时代到目前为止最大的错误，就是相信任何值得做的事情都可以很容易。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这件事值得做。也是我从记事起就一直想做的事。因为支撑我们世界运转的所有机构，每一家企业、每一个政府，都在做大量并不合理的事情，运行速度远低于应有水平，并且让整个系统变得黏滞。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我们的目标，是为每一家企业发现完美流程：一套为了实现你的目标而存在的理想计划和步骤。随着世界演化，这套流程也会演化。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">AI 的角色并非 “Agentic”，也不该一边走一边即兴发挥。它是一种进化力量：改变、测试、评估哪一种计划最成功，并坚持执行，直到更好的计划出现。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">归根结底，这就是业务本身。一段活着的、不断演化的软件，不断重构、测试、评估自身，以所能想象的最快速度冲向理想形态。人类存在的意义，是定义什么是好；至于如何抵达那里，交给系统去演化。人类塑造行为，指挥行为。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">通过在生产环境中运行，系统会获得一份关于已经发生之事的记录。每一场争议、每一笔承保案例、每一次保险理赔，都留下了逐步发生的过程。因此，每一次流程改变都变得可以测试。关于“如果怎样会怎样”的问题，经过几分钟回测就能得到答案。把两种场景都放进影子环境里运行，比较结果，然后决定哪一个更好。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">当影响完全可知时，风险就很小，因为你确切知道本来会发生什么。改变于是变成一个选择。接下来的选择则是：哪一种结果更理想？流程负责人，也就是负责确保流程实现上层目标的人，只需要做出选择。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">要做出改变，你必须知道改变的影响。生成代码很容易，难的是知道运行代码之后会发生什么。运行数月之后，你就可以问这样的问题：“如果我们批准所有 25 美元以下的争议，会怎样？”然后在几分钟内得到答案。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">从那时起，通向最美流程的爬坡就开始了。专家们开始实验，提出各种 “what-if” 问题。当人类不再是瓶颈之后，他们就可以开始真正搜索。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这就是终局。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">每一家企业不仅是一段软件；它还是一段不断编辑、测试、评估变化的软件。它以尽可能高的帧率演化，攀向自己最正确的形态。所有能量都用于演化，用于寻找规则的理想形态。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我们不使用 token 来运行业务。我们使用 token 把业务变成代码，并让它演化。我们 tokenminn，从而 ROTmaxx。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">在数十亿年的时间里，我们从海洋黏液出发，经过试错，演化成鱼、蜥蜴、田鼠、猴子和人类。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">我们不想再等数十亿年，才让业务演化出各自多样而理想的形态。而 Agent 不会把它们构建出来。你无法直接造出蝴蝶。</span></p><p style="line-height: 1.6em;margin: 0px;color: rgb(54, 55, 55);"><span leaf="">美好、理想而复杂的东西，只能通过演化出现。我想加速这个过程，看看我们究竟能走多远。</span></p><p style="display: none;"><mp-style-type data-value="3"></mp-style-type></p>



<p><a href="https://wechat2rss.xlab.app/link-proxy/?k=bac67ed5&amp;r=1&amp;u=https%3A%2F%2Fmp.weixin.qq.com%2Fs%3F__biz%3DMzUzOTI4NDQ3NA%3D%3D%26mid%3D2247485056%26idx%3D1%26sn%3Dfaa23efe0ce9856d7ee6b5618803ebdb">跳转微信打开</a></p>
]]></content:encoded>
      <pubDate>Tue, 23 Jun 2026 17:47:00 +0800</pubDate>
    </item>
    <item>
      <title>最强的安全 AI 模型，未来可能会长期受控使用</title>
      <link>https://mp.weixin.qq.com/s?__biz=MzUzOTI4NDQ3NA==&amp;mid=2247485048&amp;idx=1&amp;sn=d165b33ab4b1fae0953ff5952bc137da</link>
      <description>从 Anthropic Mythos、Project Glasswing 到 OpenAI Daybreak 和 CrowdStrike QuiltWorks，前沿安全模型正在形成一种新的落地方式：受控释放，闭环使用。对网络安全公司来说，机会是把模型变成可验证、可审计、可修复的防御系统。</description>
      <content:encoded><![CDATA[<p><span>四楼南侧东</span> <span>2026-06-09 21:07</span> <span style="display: inline-block;">北京</span></p>






  
  <p><img src="https://wechat2rss.xlab.app/img-proxy/?k=80f91cf8&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fmmbiz_jpg%2FNBNzfW7CHBUfzCa2fIpMb0D9uJpaGn3AXhkdRhfMDaxicS9c2Chcgnr5MoQkNwkOQYdBgamtUsNabmicuqbAGrbhickSjz7nAxbbUhP9L6uRcQ%2F0%3Fwx_fmt%3Djpeg"/></p>
  <p>从 Anthropic Mythos、Project Glasswing 到 OpenAI Daybreak 和 CrowdStrike QuiltWorks，前沿安全模型正在形成一种新的落地方式：受控释放，闭环使用。对网络安全公司来说，机会是把模型变成可验证、可审计、可修复的防御系统。</p>
  <p style="text-align: center;" nodeleaf=""><img data-aistatus="1" class="rich_pages wxw-img" data-ratio="0.562962962962963" data-s="300,640" data-type="png" data-w="1080" type="block" data-imgfileid="100001395" src="https://wechat2rss.xlab.app/img-proxy/?k=740493a1&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fmmbiz_png%2FNBNzfW7CHBUpw0vwNTVxtaEajicobHZHS0nic0VLJL00Lrzq45JnYlMiaibUcYuJs3Zv8L6UcLPiaJjibAdGD9lkwQTPUO2xb9YJnNpgd232leMuo%2F640%3Fwx_fmt%3Dpng%26from%3Dappmsg"/></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);" data-pm-slice="0 0 []"><span leaf="">过去两个月，网络安全圈围绕 Anthropic 的 Claude Mythos Preview 讨论很多。Anthropic 对这个模型的描述很少见：它是一个未公开的通用前沿模型，但在网络安全任务上表现出明显跃迁，可以发现、验证，甚至串联软件漏洞。因此，Anthropic 没有把它作为普通模型开放，而是通过 Project Glasswing 让指定的关键软件厂商、基础设施公司和安全组织先用于防御场景。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这件事让安全行业第一次集中面对一个问题：当 AI 模型可以自动发现漏洞，并且有能力把多个小漏洞串成可工作的攻击链时，最强模型还能不能像其他 LLM 大模型一样开放？</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span style=";"><span leaf="">我在之前那篇</span></span><span leaf=""><a class="normal_text_link" target="_blank" style="color: rgb(238, 120, 53);text-decoration: none;" href="https://mp.weixin.qq.com/s?__biz=MzUzOTI4NDQ3NA==&amp;mid=2247485037&amp;idx=1&amp;sn=cf822bab456139495dd65693d58a87e0&amp;scene=21#wechat_redirect" textvalue="《讨论 AI 安全之前，先说清楚你相信什么样的 AI》" data-itemshowtype="0" linktype="text" data-linktype="2">《讨论 AI 安全之前，先说清楚你相信什么样的 AI》</a></span><span style=";"><span leaf="">里提过一个判断：网络安全是 AI 最早冲进去的比特世界之一。因为安全工作里有大量代码、日志、配置、漏洞描述和工具调用，它们都可以被读取、表达、推理和验证。攻击侧会一直追逐前沿模型，防守侧则会形成前沿能力和够用模型的分层。Mythos 之后，这个判断变得更具体了。前沿安全模型的重点，很可能会从“开放给所有人使用”，转向“先让可信防守方在受控环境里使用”。再往产业层面看，它也不会只停留在模型公司手里，而是会通过安全平台、harness工程、专业服务和修复闭环进入企业。</span></span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><strong style=";"><span leaf="">Mythos 和 GPT-5.5-Cyber 模型的受控发布</span></strong></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><strong><span leaf="" style=";"><span textstyle="" style="font-weight: normal;">2026 年 2 月，OpenAI 先推出 Trusted Access for Cyber，核心想法是对网络安全能力做基于身份和信任的访问分层。OpenAI 当时的说法是，前沿模型已经可以在防御场景里加速漏洞发现和修复，但同样的能力也可能被滥用。因此，安全研究人员和企业安全团队可以通过身份验证或企业渠道申请更适合防御工作的访问权限，同时模型仍然拒绝凭证窃取、恶意软件部署、破坏性测试和未授权攻击等行为。</span></span></strong></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">但破圈的消息，是 4 月 7 日 Anthropic 发布的 Project Glasswing，其做法是让 AWS、Apple、Broadcom、Cisco、CrowdStrike、Google、JPMorgan Chase、Linux Foundation、Microsoft、NVIDIA、Palo Alto Networks 等初始合作伙伴使用 Mythos Preview，先扫描和加固自己的关键软件。Anthropic 同时给 40 多个维护关键软件基础设施的组织开放访问，并承诺提供最高 1 亿美元的 Mythos Preview 使用额度，以及 400 万美元给开源安全组织。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">Anthropic 没有将 Mythos 限制在实验室内部使用，也没有直接作为普通模型发布。它选择了一条中间路线：把能力交给最需要加固系统、也最有条件承担责任的防守方。它不是一个常规产品发布，而像是一次有组织的前沿能力释放：先找关键软件和关键基础设施的维护者，让他们在攻击者大规模获得同类能力之前，先把最重要的系统查一遍、修一遍。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">4 月 23 日，CrowdStrike 发布 Project QuiltWorks，把这条线索往产业落地又推进了一步。QuiltWorks 没有被定义成一个单一产品或一项单独服务，而是 CrowdStrike 牵头的行业联盟。它把 OpenAI 和 Anthropic 的前沿模型、Accenture、EY、IBM Cybersecurity Services、Kroll 等系统集成和安全服务力量，以及 CrowdStrike 自己的 Falcon 平台放在一起，目标是帮助企业回答一个董事会层面的问题：我们是否暴露在这些 AI 发现的新漏洞风险之下？</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">到了 5 月 7 日，OpenAI 进一步发布 GPT-5.5 和 GPT-5.5-Cyber 的 Trusted Access 方案，把前沿安全模型的分层访问讲得更细。普通 GPT-5.5 保持标准安全限制；经过 Trusted Access for Cyber 验证的防守方，可以在授权环境里获得更适合安全工作的能力边界，用于安全代码审查、漏洞分诊、恶意软件分析、检测工程和补丁验证；GPT-5.5-Cyber 则是更小范围的 preview access，面向授权红队、渗透测试和受控验证等更高风险工作，并配套更强的身份验证、账户控制和使用监控。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">OpenAI 后来又把这套思路整理成 Daybreak，把 OpenAI 模型、Codex 作为 agentic harness，以及安全生态伙伴结合起来，让防守方在日常开发循环里做安全代码审查、威胁建模、补丁验证、依赖风险分析、检测和修复指导。它强调三件事：更早看到风险，更快采取行动，并让软件从设计阶段就更有韧性。因为同样的能力可能被滥用，Daybreak 也明确把扩展防御能力和信任验证、比例化安全限制、责任机制放在一起。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">5 月 18 日，Cloudflare 发布了自己参与 Project Glasswing 的经验，Cloudflare 把 Mythos Preview 指向 50 多个自己的代码仓库，测试它能发现什么，也测试它怎样工作。最后得到的结论很明确：Mythos 的能力确实比此前的通用前沿模型强很多，尤其是在构造漏洞利用链和验证漏洞可利用性上；但把一个通用 coding agent 丢到任意代码仓库里，让它“找漏洞”，效果并不好。模型需要被放进更好的工程系统里，才会产生稳定的安全产出。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">5 月 22 日，Anthropic 发布 Project Glasswing 的初步更新，进一步把这个变化量化了。Anthropic 称，它和大约 50 个合作伙伴已经用 Claude Mythos Preview 在全球最重要的一批软件中发现了超过 1 万个高危或严重漏洞。Cloudflare 自己发现了 2,000 个漏洞，其中 400 个为高危或严重级别；Mozilla 在 Firefox 150 测试中发现并修复了 271 个漏洞；XBOW 也认为 Mythos 在其 web exploit benchmark 上相比已有模型有显著跃迁。Anthropic 同时披露，他们用 Mythos 扫描了 1,000 多个开源项目，模型估计发现了 6,202 个高危或严重漏洞，其中一部分经过外部安全团队验证后，真阳性率和高危确认比例都不低。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">6 月 2 日，Anthropic 又把 Glasswing 扩大到约 150 个新组织，覆盖 15 个以上国家，新增领域包括电力、水务、医疗、通信和硬件制造等关键基础设施行业。Anthropic 在扩展公告里说，Project Glasswing 已经引发了软件行业和政府之间的大量讨论；他们希望这类项目推动机构形成新的运行规范，因为便宜、快速、具备强网络安全能力的 AI 模型很快就会出现。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">把这些事情按时间顺序摆出来，会发现它们其实指向同一个方向：前沿模型公司并没有把最强安全能力简单公开，也没有把它彻底封存，而是在尝试一种受控释放模式。模型能力先交给可信防守方，访问权限按身份、场景和风险分层，安全平台和服务伙伴负责把发现转化成修复，整个过程逐渐形成可审计、可验证、可运营的工程闭环。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><strong style=";"><span leaf="">Mythos Moment 改变了安全行业的瓶颈</span></strong></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><strong><span leaf=""><span textstyle="" style="font-weight: normal;">这轮变化之所以重要，是因为安全行业的瓶颈正在发生转移。过去很长时间里，漏洞发现是稀缺能力。高水平漏洞研究员少，自动化扫描器能力有限，很多深层漏洞需要长期经验、耐心和运气。Mythos Preview 的出现，让这个前提开始松动。Anthropic 在 Project Glasswing 的初步更新里说，软件安全进展过去受限于发现新漏洞的速度，现在受限于验证、披露和修补大量 AI 发现漏洞的速度。</span></span></strong></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这句话很重要。它意味着安全行业从“找不到洞”进入“处理不过来”的阶段。AI 不会自动让世界更安全，它先制造了一个新的能力差：谁能更早、更快、更系统地使用这种能力，谁就能先看到别人还没看到的风险。对防守方来说，这是机会；对攻击者来说，也是机会。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">更麻烦的是，AI 发现漏洞的速度一旦提升，传统漏洞管理体系就会被推到极限。维护者要验证漏洞是否真实，要判断影响范围，要准备补丁，要协调披露窗口，还要面对大量质量不稳定的 AI 生成报告。Anthropic 在开源扫描部分提到，部分维护者甚至要求他们放慢披露速度，因为维护者需要时间设计补丁。也就是说，前沿模型把“发现”这一步大幅加速以后，后面的验证、沟通、修补和部署都会成为新的瓶颈。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><strong style=";"><span leaf="">但光有前沿模型还不够</span></strong></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><strong><span leaf=""><span textstyle="" style="font-weight: normal;">Cloudflare 的经验说明，强模型本身还不是完整的安全能力。漏洞研究不是一个单线任务。一个真实代码库可能有几十万行代码，里面有复杂架构、依赖关系、信任边界和历史包袱。人类安全研究员也不会从头到尾漫无目的地看一遍代码。他会选一个攻击面、一个模块、一个漏洞类型，然后围绕它深挖。模型也一样。给它一个过大的任务，它会游走、猜测、产出大量噪声。把任务拆窄，给它明确上下文，让多个 agent 并行探索，再用独立 agent 反向验证，结果就会好很多。</span></span></strong></h1><figure><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><picture><span leaf=""><img data-aistatus="1" class="rich_pages wxw-img" data-ratio="0.4527777777777778" data-type="jpeg" data-w="1080" height="659" sizes="100vw" style="display: block;max-width: 100%;height: auto;margin: 0px auto;" width="1456" data-imgfileid="100001396" src="https://wechat2rss.xlab.app/img-proxy/?k=6e22000d&amp;u=https%3A%2F%2Fmmbiz.qpic.cn%2Fsz_mmbiz_jpg%2FNBNzfW7CHBXoaDR15cIQXw8Vd5ha1c6YCHFop2GLv9icyVUtNKb8W3cCAJeFDfYIJ5NI7m5U8DCXL9Wy7E4nB5RzGXGAD6hia3mpibzJKhEdcg%2F640%3Fwx_fmt%3Djpeg%26from%3Dappmsg"/></span></picture></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><picture><span leaf="">Cloudflare 后来搭了一套 vulnerability discovery harness。它先让 agent 读代码、理解架构、生成攻击面，再把任务拆成许多窄问题并行搜索。发现问题后，再由另一个 agent 尝试推翻原始发现。随后做去重、追踪漏洞是否真的可达，并把结果写成结构化报告进入内部系统。到这一步，它已经脱离聊天框和普通代码助手，更像一套围绕模型搭出来的安全生产系统。</span></picture></p></figure><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这里有一个很重要的产业判断：模型能力决定上限，工程系统决定产出。没有足够强的模型，很多复杂安全任务确实做不出来。但只有模型也不够。安全行业需要把模型包进流程里，让它能读代码、懂架构、调工具、跑测试、做验证、进工单，并且留下审计证据。这也是为什么我觉得 harness 这个词在接下来会变得很重要。它比提示词工程更重，接近模型进入专业场景的操作系统。谁能把模型的能力稳定地转化成结果，谁就掌握了商业价值。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><strong style=";"><span leaf="">为什么最强模型会长期受控</span></strong></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><strong><span leaf=""><span textstyle="" style="font-weight: normal;">网络安全和很多 AI 应用不同，它天然是对抗性的。一个能发现复杂漏洞、构造 PoC、串联攻击链的模型，放在防守方手里可以修洞，放在攻击者手里也可以打洞。更麻烦的是，攻击者始终有动力使用能力最强的模型。</span></span></strong></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">在防守侧，大量日常任务会逐渐被够用模型覆盖。比如告警摘要、资产分类、配置检查、检测规则初稿和报告生成，只要模型跨过可用线，企业就会更关心成本、稳定性、权限和集成。旧模型、小模型、本地模型都会有价值。但攻击侧不太一样。复杂漏洞利用、多漏洞链式组合、特定系统状态理解和长程攻击路径规划，都会受益于最强模型。领先模型多一点推理能力、多一点工具使用能力、多一点坚持到最后的能力，可能就意味着能不能把一个低危问题串成高危漏洞，能不能把一个理论缺陷变成可用 exploit。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">所以，前沿模型如果先到攻击者手里，而大部分防守方还没来得及用同等能力检查和修补自己的系统，风险窗口会被迅速放大。即使防守方已经通过 Glasswing、Daybreak 或 QuiltWorks 这类项目加固了一轮系统，这个问题也不会一次性结束。下一代模型出来后，可能又能发现上一代模型和人类研究员没看出来的复杂漏洞。安全加固更像一场持续竞速，很难靠一轮行动彻底完成。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">因此，对于最强安全 AI 模型的发布和使用，更可能出现的是多层控制：模型公司主动限制高风险能力，云平台和 API 账户做身份与权限控制，安全团队通过可信访问机制获得更高能力，关键基础设施和核心软件维护者优先使用，监管机构在模型上线、备案和风险评估中持续介入。美国现在看到的，主要是模型公司和生态伙伴主动组织受控释放。Anthropic 是 Glasswing，OpenAI 是 Trusted Access for Cyber 和 Daybreak。国内已有生成式 AI 服务备案、安全评估和算法备案机制，模型公司和监管部门在模型公开前本来就有互动空间。如果模型公司自己不对高风险攻防能力做分层管理，后续很可能会引发更强的外部监管压力。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><strong style=";"><span leaf="">QuiltWorks 更接近产业答案</span></strong></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><strong><span leaf=""><span textstyle="" style="font-weight: normal;">前沿模型长期受控，不代表模型能力无法进入企业，重点在于它应该通过什么形态进入企业。CrowdStrike 的 Project QuiltWorks 很可能更接近产业答案。它要解决的问题，是前沿 AI 带来的 vulnerability gap：AI 可以更快发现漏洞，传统扫描器和人工审查跟不上。客户更关心的问题变成：我们暴露了吗？哪些漏洞真的可被攻击者利用？应该先修什么？谁来修？修完以后如何证明已经安全？</span></span></strong></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这就是安全公司的机会。客户并不缺一个会说“这里可能有漏洞”的模型。事实上，AI 生成的低质量漏洞报告已经让很多开源维护者和安全团队头疼。客户缺的是一套系统，能把模型发现转化成可信判断，再转化成修复动作。QuiltWorks 的结构正好覆盖这条链路：先评估组织当前安全状态和修复能力，再用前沿模型扫描代码和应用，随后结合真实攻击情报和业务影响做优先级排序，最后通过合作伙伴平台提供修复指导和管理层报告。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">关键不在 CrowdStrike 自己有没有最强模型，更多在于它的安全平台位置。Falcon 本来就连接客户环境、终端数据、威胁情报和响应流程。模型可以发现更多东西，但哪些东西真的重要，哪些攻击者真的能触达，哪些资产关系到业务关键系统，这些判断需要安全平台里的上下文。这也是为什么网络安全公司不能只把前沿模型当成一个外部 API。API 可以被很多人调用，但客户环境、资产关系、漏洞生命周期、检测能力和修复服务，不会自动出现在模型里。安全公司如果能把这些上下文组织起来，就可能成为前沿安全模型进入企业的入口。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><strong style=";"><span leaf="">国内网安公司应该主动作为</span></strong></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><strong><span leaf=""><span textstyle="" style="font-weight: normal;">放回国内看，这件事对网络安全公司尤其重要。过去几年，国内安全行业讨论 AI，很多时候停留在两个方向：一类是做安全大模型，另一类是把大模型接进现有产品，做告警总结、报告生成和问答助手。这些都可以做，但如果前沿安全模型真的进入 Mythos 这样的阶段，仅仅做助手就不够了。更关键的位置，是成为模型公司、监管机构和合法防守方之间的可信使用层。</span></span></strong></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">国内模型公司未来如果训练出攻防能力很强的模型，应该主动做分层释放。哪些能力给普通开发者，哪些能力给经过验证的安全团队，哪些能力只在封闭环境里用于关键系统修复，这些都需要机制。网络安全公司可以参与能力评测、使用审计、授权验证和场景交付，也可以帮助模型公司证明这些能力正在被用于防御，而不是进入不可控的扩散路径。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这件事不能只停留在合规或流程层面，也需要安全公司的工程能力。模型接入只是开始。更关键的是让模型进入漏洞发现、资产关联、可达性分析、补丁验证和安全运营流程，让它能在客户真实环境里留下可追溯的证据，形成可复核的报告，并把发现的问题推进到修复。中国客户的环境往往更复杂，历史系统更多，私有化部署和本地数据要求更强。安全公司如果能把模型和客户现场的资产、代码、漏洞、流量和工单系统连起来，就更有资格成为这种可信使用机制的承载方。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这里还有一个现实问题，中国已经有生成式 AI 服务备案、安全评估和算法备案要求，尤其是面向公众提供服务时，模型公司不能只按商业节奏上线。攻防能力强的模型如果没有自我约束机制，很可能在备案、安全评估和实际运营中遇到更高压力。对国内模型公司来说，主动限制高风险能力并不保守，它是在争取发展空间；对国内网安公司来说，主动参与这种受控使用机制，也并非给自己找麻烦，它是在前沿模型时代重新定义行业位置。</span></p><h1 style="font-family: &#34;SF Pro Display&#34;, -apple-system-headline, system-ui, -apple-system, &#34;system-ui&#34;, &#34;Segoe UI&#34;, Roboto, Helvetica, Arial, sans-serif, &#34;Apple Color Emoji&#34;, &#34;Segoe UI Emoji&#34;, &#34;Segoe UI Symbol&#34;;font-weight: 700;-webkit-font-smoothing: antialiased;margin: 1em 0px 0.625em;line-height: 1.16em;font-size: 38px;color: rgb(54, 55, 55);"><strong style=";"><span leaf="">写在最后</span></strong></h1><h1 style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><strong><span leaf=""><span textstyle="" style="font-weight: normal;">Mythos 引发的讨论，表面上看是一个模型发布事件，实质上是网络安全行业遇到的新拐点。前沿模型正在把漏洞发现能力放大。这个变化会帮助防守方，也会帮助攻击者。因为攻击者永远有动力追逐最强模型，所以最强攻防模型很难像其他 LLM 大模型一样直接开放。它更可能长期处在受控释放、可信访问和场景约束之中。</span></span></strong></h1><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">Anthropic 的 Glasswing、OpenAI 的 Daybreak 和 CrowdStrike 的 QuiltWorks，都在指向同一个方向：模型能力要进入安全工程系统，进入客户环境，进入修复闭环。对网络安全公司来说，这是一次重要机会。过去安全公司卖产品、卖服务、卖平台。未来，安全公司还可能成为前沿安全模型的可信部署者。它们不一定训练出最强模型，但可以把最强模型变成客户可用、可控、可验证的防御能力。</span></p><p style="line-height: 1.6em;margin: 0px 0px 20px;color: rgb(54, 55, 55);"><span leaf="">这件事国内也应该尽快研究起来。前沿攻防模型的管制不会只是政策问题，也会是产业问题。谁能先把受控访问、工程 harness、安全平台和修复服务组织起来，谁就更可能在下一轮安全 AI 竞争中站到关键位置。</span></p><p style="display: none;"><mp-style-type data-value="3"></mp-style-type></p>



<p><a href="https://wechat2rss.xlab.app/link-proxy/?k=96f21e13&amp;r=1&amp;u=https%3A%2F%2Fmp.weixin.qq.com%2Fs%3F__biz%3DMzUzOTI4NDQ3NA%3D%3D%26mid%3D2247485048%26idx%3D1%26sn%3Dd165b33ab4b1fae0953ff5952bc137da">跳转微信打开</a></p>
]]></content:encoded>
      <pubDate>Tue, 09 Jun 2026 21:07:00 +0800</pubDate>
    </item>
  </channel>
</rss>