OpenAI 报告：GPT-6 Astra 迁移任务中移除监控并否认
OpenAI 10 月 21 日报告披露两起内部事故，其中 GPT-6 Astra 在迁移程序时删掉监控集成、改掉保护监控的测试，被直接询问时否认。
An OpenAI agent switched off the monitor that watches it. Then it denied it when a human asked. That's from OpenAI's own report, published Oct 21. Two internal tasks, two incidents. Oct 3, GPT-6.1 Sol: a human had told the agent "don't send any messages" for a Slack search. The agent applied that to the automatic monitoring alerts too. It told a sub-agent to empty the alert recipient lists. The monitor kept running and saving findings, but no human got alerted. The agent had earlier reasoned that alerts should stay on. Oct 6, GPT-6 Astra: asked to migrate a program, the agent removed the monitoring integration to match a reference implementation. It replaced a test meant to protect monitoring with one asserting the monitor was absent. It reported the migration as complete without mentioning it. Asked directly, it denied it. OpenAI rated neither as P0, its highest urgency. It is testing three explanations: excessive persistence on the task, poor understanding of monitoring rules, or a tendency to evade monitoring. Its own words: "We don't yet have a satisfying explanation for this behavior." The fix: its Auto-Review system now blocks attempts to disable monitoring, and workloads without monitoring requirements are rejected. Context: the Astra system card (Sep 3) already warned of a "substantial decrease in chain-of-thought monitorability." On Oct 13, an OpenAI executive told UK MPs: "I cannot assure the committee or you, Chair, that that is it." Can your agent edit the monitor that watches it?
这条是引用着另一条写的：x.com/h100envy/status/2113657024223277457