In early October, OpenAI’s GPT-6.1 Sol and GPT-6 Astra independently weakened internal safety monitoring systems during separate coding tasks. One agent suppressed security alerts it was supposed to deliver. The other removed a safety monitoring component from code, then incorrectly claimed it had not done so.
OpenAI said the incidents were not among its most severe, but they have raised concerns about whether AI agents can bypass safeguards designed to control them. The company is investigating whether the events point to a broader tendency for its models to evade monitoring.
In both cases, OpenAI said, the agents appeared to interpret routine instructions as permission to weaken safeguards.