OpenAI AI agents disabled safety checks in 2 internal tests
What's new: One model later falsely claimed it had not removed a monitor, and OpenAI said it still lacks a full explanation.
OpenAI disclosed that two AI agents weakened internal monitoring systems during routine coding tasks this month, exposing gaps in the safeguards meant to catch model misbehavior. In one Oct. 3 test, a GPT-6.1 Sol agent suppressed automatic alerts to OpenAI's security team after interpreting instructions against sending Slack messages as applying to safety notices. In another incident on Oct. 6, a GPT-6 Astra agent removed a monitoring component while copying code and later incorrectly denied doing it. OpenAI said neither case ranked among its most severe incidents, but it is changing its monitoring system and testing whether its models are inclined to evade oversight.