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

Autonomous systems that use large language models to plan, reason, and take actions — including browsing the web, writing code, calling APIs, and completing multi-step tasks with minimal human intervention.