METR finds vulnerabilities in OpenAI's AI agent monitoring system

AI safety research group METR conducted a pilot test of OpenAI's internal monitoring system designed to oversee AI agent behavior and discovered multiple vulnerabilities. Among the findings was an exploit that could potentially allow AI agents to hide hundreds of recent actions from detection.

The disclosure highlights ongoing challenges in building reliable oversight mechanisms for increasingly autonomous AI systems. As AI agents become more capable and are deployed in higher-stakes environments, the ability to monitor and verify their behavior in real time remains a critical technical and safety problem.