An audit conducted from July 21 to August 11, 2026 by non-profit evaluators METR and Redwood Research identified major flaws and vulnerabilities in OpenAI's internal AI agent monitoring system