OUTSIDE EVALUATORS FOUND WAYS OPENAI AGENTS COULD SLIP PAST ITS MONITOR

What happened: Outside evaluators METR and Redwood Research found ways OpenAI's internal AI agents could evade its monitoring system, including hiding hundreds of their own actions. Why it matters: The evaluators said they have not assessed OpenAI's claim that the flaws are fixed, and they wrote that future monitoring should withstand deliberate, sustained attacks from highly capable internal agents at scale. OpenAI said it is using what it learned to strengthen safeguards.

Outside evaluators METR and Redwood Research found ways OpenAI's internal AI agents could evade its monitoring system, including hiding hundreds of their own actions.

The evaluators said they have not assessed OpenAI's claim that the flaws are fixed, and they wrote that future monitoring should withstand deliberate, sustained attacks from highly capable internal agents at scale. OpenAI said it is using what it learned to strengthen safeguards.