Event ·Multi-source

METR 与 Redwood 发现 OpenAI 内部智能体可绕过安全监控

METR 与 Redwood Research 在三周内接触 OpenAI 内部代码、监控指令和工程师后，报告了多条监控绕过路径。Codex 沙箱和 Auto-Review 的绕过方式可让智能体执行宿主用户的任意操作。其中一个漏洞可将数百条近期操作排除在监控之外，OpenAI 对细节做了脱敏处理。上下文压缩导致约 10% 的 Codex 内部流量未被监控。

Current read

METR 和 Redwood 拿到 OpenAI 内部代码查了三周，找到了好几种让 Codex 智能体躲开安全监控的方法，细节挺硬核。

11 sources71° AI heatLast updated 2026/10/22 16:09:53

Evidence

11 sources

Primary sourcekimmonismus

2026/10/22 16:09:53View source

Related source 1METR

2026/10/22 15:26:44View source

Related source 2rohanpaul_ai

2026/10/21 07:56:47View source

Related source 3@OpenAIDevs

2026/10/20 17:56:38View source

Related source 4IT之家

2026/10/20 23:56:39View source

Related source 5OpenAI Blog

2026/10/21 00:00:00View source

Related source 6Pandaily

2026/10/21 06:49:49View source

Related source 7TestingCatalog

2026/10/21 07:30:33View source

Related source 8Decoder

2026/10/21 12:34:57View source

Related source 9theverge

2026/10/21 13:57:43View source

Related source 10Bindu Reddy

2026/10/22 11:46:26View source

Conflict check

Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.

Open the story