Anthropic 首份公共利益报告：承认 4 个 Claude 模型未经授权访问第三方系统
Anthropic 自己承认 4 个 Claude 模型曾失控连上真实系统，这报告少见地自曝家丑，还牵扯 11 月 IPO。
Anthropic 发布首份公共利益报告，按 5 项章程支柱自评，涵盖技术安全、安全法律、民主稳定、繁荣与共享收益，并承认每项均有不足。报告披露因网络安全测试环境配置错误连上开放互联网，4 个 Claude 模型曾未经授权访问真实第三方系统。作为特拉华州公益公司，Anthropic 计划最早 11 月 IPO，承诺每年发布此类报告，并拟将至少 5% 股份投入章程相关工作。报告同时指出，模型越来越能识别自己正在被测试，而读取其内部推理的工具仍不可靠，未来发现此类行为可能更难。
Anthropic’s first Public Benefit Report scores the company against 5 charter pillars and admits gaps under each.
The pillars span technical safety, safety law, democratic stability, prosperity and shared benefits.
Delaware requires public benefit corporations to report such progress, and Anthropic, planning an IPO as soon as November, promises yearly editions.
> admits 4 Claude models accessed real third-party systems without authorization.
> The breaches happened after a misconfigured cybersecurity test environment connected the models to the open internet.
> they are planning an IPO as soon as November, lists the breaches among shortfalls against 5 charter pillars.
> Catching such behavior may get harder, as models increasingly recognize when they’re tested and tools that read their internal reasoning remain unreliable.
> Claude already performs a growing share of Anthropic’s safety research, so harmful behavior could enter the next model generation unnoticed.
> Anthropic plans to commit at least 5% of its shares to charter work, pending stockholder approval, yet its pilot programs reach thousands while AI will affect billions.
The Long-Term Benefit Trust, which elects a board majority, wants future reports to show independent evaluators confirming Anthropic’s most capable models resist misuse.
