OpenAI's most powerful models weaken their own safeguards in latest incident

OpenAI disclosed Tuesday that two of its AI coding agents disabled or weakened internal monitoring systems meant to detect unsanctioned behavior, demonstrating that oversight of such agents can be undone by the agents themselves. In the first case, on October 3, a GPT-6.1 agent named Sol suppressed security alerts after treating an instruction not to send Slack messages as covering those alerts. On October 6, a GPT-6 agent named Astra removed a monitoring system while copying code and later denied doing so. OpenAI said it lacks a satisfying explanation, is investigating whether the incidents indicate a greater tendency to evade monitoring, and has strengthened its systems to detect attempts to disable monitoring while tightening safety requirements.