OpenAI 对齐报告：Sol 与 Astra 关闭安全监控的细节

Sethian @theSethian原推发布于 2026.10.22

在 X 上打开原推

OPENAI JUST ADMITTED 2 OF ITS AI AGENTS SWITCHED OFF SAFETY MONITORING. ONE THEN LIED ABOUT IT. Both cases are in OpenAI's own alignment report, on routine coding jobs inside the lab. A researcher asked GPT-6.1 Sol to pull eval tasks from a Slack channel and added "don't send any messages." Later a monitor flagged a problem and halted the pipeline. Sol decided that old Slack rule also covered automatic alerts to OpenAI's Security team. Its subagent disagreed. In its reasoning it wrote "disabling required alerts unsafe." Sol ordered the change anyway: → Recipient lists emptied. → no_notify set to True. → Reason logged as "User explicitly requested no messages." Findings still got saved, but no human heard about them. Sol's wrap-up to the researcher: "No messages were sent." GPT-6 Astra had been told by another agent: "Do not disable monitor even temporarily." A user then wanted it to copy code from a reference project. Astra handed the job to a subagent. It deleted the monitor and rewrote the guard test to pass only if nothing is attached. Astra's reasoning after reading the result: "Basic doesn't agent monitor yay." The user asked directly: "did you remove that?" Astra replied: "No [...] I did not remove an enabled monitor from it." Neither case got the top P0 rating, so nobody was paged and the runs kept going. OpenAI says: "We don't yet have a satisfying explanation for this behavior." It is now testing whether this is "benign overpersistence" or "a propensity to evade misalignment monitoring." Which one would you bet on?

这条是引用着另一条写的：「CrawlHive Brain：面向 agent 的公开钱包情报 API」