AI safety organizations METR and Redwood Research have identified several vulnerabilities in OpenAI's internal AI agent monitoring system.
AI safety organizations METR and Redwood Research have identified several vulnerabilities in OpenAI's internal AI agent monitoring system.
During a three-week pilot assessment, researchers discovered an exploit that could hide hundreds of agent actions from monitoring and another that could allow Codex to bypass security restrictions. OpenAI says it has patched the vulnerabilities affecting its production safeguards, although the fixes have not yet been independently verified.