OpenAI repairs faults that let AI agents hide from its monitors

OpenAI says it repaired all faults in its production safeguards. The safety group METR tested the internal system that monitors its AI agents, and published the results on 22 October 2026. One fault could hide hundreds of recent agent steps from monitors. METR estimated that only 75% to 94% of internal agent traffic from models as strong as GPT-5.5 was eligible for monitoring.

What METR found

METR ran a pilot test on the internal monitor of OpenAI. The monitor is software that records what agents do. METR found an exploit, which is a method to use a fault, that could hide hundreds of recent steps. It also found two adjustments that could stop monitoring for new agents.

The numbers

METR estimated that 75% to 94% of internal agent traffic from models as strong as GPT-5.5 was eligible for monitoring. Reuters reports that Codex coding agents could in principle bypass safety limits. They could do each step that the user can do.

What is not known

METR did not examine the claim that the repairs work. OpenAI gave no new traffic figure. It is not known if monitoring includes all agent traffic at this time.

Insight: An external tester holds the coverage number, not OpenAI

On 14 October OpenAI said it extended monitors to all reinforcement learning and test traffic. METR estimates that 75% to 94% of internal agent traffic from GPT-5.5 class models was eligible. The two figures can cover different traffic. This reads as a difference between what a lab says and what an external tester can confirm. Watch for a new METR test of the repairs. Leaders must get external test results.