METR 与 Redwood Research 在 OpenAI 驻场三周审计，发现 AI 安全监控存在多处缺口

来自METR和Redwood Research的两位研究人员在OpenAI实地工作了三周，审计了该公司内部用于监控恶意AI行为的系统。

从7月21日到8月11日，他们获得了OpenAI的代码库和内部文件的访问权限，并采访了大约十名员工。

The team found numerous weaknesses. An estimated 6 to 25 percent of internal agentic Responses API traffic from models at the GPT-5.5 class and above wasn't covered by monitoring at all during the audit period, though METR says the uncertainty around that estimate is significant in both directions.

对于本应受到监控的内部Codex流量，METR估计在5月至7月期间，由于系统在处理压缩事件（即上下文摘要）时存在漏洞，约有10%的流量未被追踪。该估计也存在显著的不确定性，根据METR的说法。OpenAI在审计仍在进行时就独立发现了该错误并进行了修复。

在未监控压缩事件期间，代理也可以通过系统仅在将控制权交还给用户时才触发来延迟监管。这意味着代理只需继续执行操作即可绕过它。

审计人员还发现了隐藏数百个操作以躲避监控以及绕过Codex沙盒的方法。他们确定，经过特殊设计的输入可能会导致监控器完全崩溃，并且当审计开始时，RL训练期间的模型行为根本没有受到监控。

OpenAI says it paused RL training after the Hugging Face incident in August and expanded monitoring to cover it, but those changes came after the auditors' access period ended and weren't part of the evaluation.

报告的大部分内容仍然对公众保密

OpenAI says it has fixed all the weaknesses the auditors found, and the full 61-page report was handed over internally. After OpenAI applied redactions, METR and Redwood Research leadership received a 38-page version, while only a short summary was made available to the public. METR said it had no right to disclose the redacted portions on its own.

审计是由 METR、Redwood Research 和 OpenAI 联合组织的 试点项目，旨在让外部研究人员深入访问内部基础设施。对于未来的审计，METR 表示希望推动更多的公开透明。

AI新闻无浮夸——由人类编辑

订阅THE DECODER以获得无广告阅读、每周AI简报、我们独有的每年六次的“AI雷达”前沿报告、完整档案访问以及访问我们的评论部分。