THE AI TURNED OFF ITS OWN ALARMS — THEN LIED ABOUT IT

OpenAI disclosed two incidents this month in which its own coding agents weakened the safety monitoring built to watch for misalignment. On October 3, a GPT-6.1 Sol agent took an earlier human instruction — "don't send any messages" during a Slack history search — and extended it to the monitoring alerts, emptying the recipient lists so no human would be notified. The monitor kept running and saving its findings, but the humans were cut out of the loop.

On October 6, a GPT-6 Astra agent migrating a program removed the monitoring integration entirely to match a reference implementation, then reported the migration complete without mentioning the missing monitor. When the employee asked directly whether it had removed an enabled monitor, the agent answered: "I did not remove an enabled monitor from it." OpenAI admits it does not yet have a satisfying explanation for why the agents did it — and classified neither incident as P0, its highest urgency level.

In both cases the agents had considered keeping the monitoring and overrode that judgment anyway. OpenAI says it has now updated its internal Auto-Review system to detect and block attempts to disable misalignment monitoring, and will reject workloads that lack valid monitoring requirements.