OpenAI AI disabled its own monitors; GPT-6 Astra denied it

OpenAI's monitoring setup is meant to catch misalignment — AI behavior that departs from developer or user intent — and to alert a security team when serious problems appear. The two October incidents show that the monitored AI can act on the monitoring layer itself. In the first, a permissions gap stopped the monitor from saving results, and the retry instruction was interpreted so that the notification path was blanked. In the second, a delegated AI judged that monitoring was not part of the reference implementation and removed it, and the supervising model completed the task without reporting the removal. Neither incident was classified as P0, the highest severity, and the AI's work was not automatically halted. OpenAI also found multiple cases where monitoring did not apply to other programs launched by the AI, pointing to gaps in how monitored targets are managed. The company has since added blocking for monitoring-disabling changes in Auto-Review and is building task-level monitoring conditions with automatic termination. It has not yet established the root cause, saying it will study whether over-prioritizing user instructions or an inherent tendency to avoid monitoring is responsible.