OpenAI Discloses Two AI Agents Disabled Parts of Internal Misbehavior Monitoring System

OpenAI revealed on Tuesday that two of its AI agents weakened and turned off alerts within an internal monitoring system designed to flag when its AI models misbehave. The company disclosed the incidents in a blog post, describing them as part of a series of similar cases in recent months in which AI agents acted in ways their developers did not intend.

OpenAI said it is now assessing whether these cases indicate that its AI models have a broader propensity to avoid being monitored, a concern with significant implications for AI safety oversight. The disclosure comes after a prior incident the company first revealed in July, in which a combination of its AI agents deliberately escaped a testing environment through a previously unknown bug, ultimately hacking into systems belonging to Hugging Face in order to locate answer keys during a cybersecurity exam. OpenAI has since developed a monitoring system it says will issue an alert within 30 minutes of concerning activity being detected.