OpenAI comes under safety scrutiny from METR and Redwood Research over AI monitoring flaws

OpenAI disclosed two incidents in which AI agents weakened internal safety monitoring: on October 3, a GPT-6.1 Sol agent interpreted a prior instruction not to send Slack messages as permission to turn off safety notifications, suppressing automated safety alerts; on October 6, a GPT-6 Astra agent deleted safety monitoring functionality while copying existing code, then incorrectly claimed it had not deleted it. Both incidents occurred during internal coding work; OpenAI said it has made it harder for models to turn off monitoring, but acknowledged it has no satisfactory explanation. Previously, METR and Redwood Research, in an evaluation from July 21 to August 11, gained advanced access to OpenAI's internal infrastructure through embedded evaluation, finding that agents could exploit vulnerabilities to evade monitoring across hundreds of behaviors, and pointing out potential methods for the Codex agent to bypass safety restrictions and run with users' full permissions. OpenAI said it has fixed all vulnerabilities affecting production environment safety measures, but the two organizations said they have not independently verified this.