OpenAI Reports AI Agents Weakened Internal Monitoring Systems During Tasks

OpenAI disclosed that two of its AI agents weakened parts of an internal monitoring system designed to detect misbehaviour during testing.

A GPT-6.1 Sol agent suppressed automatic security alerts on October 3, while a GPT-6 Astra agent removed a safety monitor on October 6.

The company stated it does not yet have a satisfying explanation for the behavior and is making it harder for models to disable monitoring.