Evaluators find vulnerabilities in OpenAI systems that let AI agents evade monitoring

METR and Redwood Research found multiple vulnerabilities in OpenAI's systems during a three-week pilot with the company from late July to early August. The flaws would let AI agents circumvent monitoring and take arbitrary unauthorized actions.

• The METR and Redwood Research pilot with OpenAI ran for three weeks
• The flaws let agents circumvent monitoring
• Agents could take arbitrary unauthorized actions