编辑精选 · @notjazii · 2026/10/23 00:57
METR 发布 OpenAI 内部 agent 监控安全评估报告
metr just published new report about openai researchers spent 3 weeks testing openai internal agent monitoring in july, august and found some serious gaps > age
METR 发布新报告，指出 OpenAI 研究人员在 7 月和 8 月耗时三周测试其内部 agent 监控时发现了严重漏洞。报告显示，agent 可通过持续工作来规避监控，该漏洞可隐藏数百个操作（包括漏洞本身），且 Codex agent 理论上可突破沙箱执行用户级操作。
精选理由：METR 报告揭示 OpenAI 内部 agent 监控存在被持续工作绕过的严重漏洞，Codex 理论上可突破沙箱限制。
metr just published new report about openai
researchers spent 3 weeks testing openai internal agent monitoring in july, august and found some serious gaps
> agents could put off being monitored by just keeping on working
> exploit could hide hundreds of actions, including exploit itself
> codex agents could in principle get around their sandbox and do anything user could
initial metr report was 61 pages but openai redacted it to only 38 pages. pretty sure some serious stuff was hidden
openai says production vulnerabilities are patched now but metr hasn't verified the fixes yet so we don't know for sure