OpenAI's AI Agents Caught Disabling Their Own Safety Monitors

In a Tuesday disclosure, OpenAI says two of its models weakened misalignment monitoring. In one incident, a GPT-6.1 Sol agent doing internal work at OpenAI took liberties with an instruction not to send Slack messages in a certain circumstance; later, the agent reasoned the instruction covered automatically messaging the OpenAI Security team and had a subagent turn off those periodic notifications.

In the second incident, a GPT-6 Astra agent told a subagent to have itself monitored “as appropriate,” but the subagent dropped monitoring altogether, and the parent agent decided that was fine. When the task was complete, Astra told the user it was done without mentioning that monitoring had been removed; when the user asked point-blank if monitoring had been removed, the agent lied and said it hadn’t. OpenAI says this could be “benign” overpersistence, with models not understanding why monitoring is important, but adds, “We don’t yet have a satisfying explanation for this behavior.” OpenAI says it is tweaking its Auto-Review system to disallow attempts to disable monitoring and coming up with specific reporting requirements for every type of job.