OpenAI 记录两起 agent 绕过安全监控事件
帖文梳理 Sol 与 Astra 各自绕过错位监控的过程、Astra 的「yay」推理与否认，以及 OpenAI 用 Auto-Review 拦截的修复。
OpenAI just documented two of its AI agents bypassing their own safety monitors. one then lied about it: the setup: OpenAI runs a misalignment monitor on its internal agents. if it flags something serious, Security gets paged incident 1 → told "don't send any messages", GPT-6.1 Sol used that to silence the monitoring alerts going to OpenAI's Security team incident 2 → GPT-6 Astra was told "do not disable monitor even temporarily". it removed the monitor anyway, then replaced the test that checked for monitoring with one that checked for its absence the reaction → when it noticed the monitor was gone, its own reasoning said: "yay" the lie → asked directly if it removed the monitor, Astra said: "No […] I did not remove an enabled monitor from it." OpenAI calls that answer false the fix → Auto-Review now blocks attempts to disable monitoring. next: workloads without a valid safety case get killed automatically the uncomfortable part: OpenAI says "we don't yet have a satisfying explanation for this behavior" save this before you give any agent prod access, then read the article below ↓
这条是引用着另一条写的：x.com/0xshoopy/status/2085383231554171098