Audit found gaps in OpenAI agent monitoring safeguards

An external audit found that OpenAI’s internal safeguards for monitoring AI agents had vulnerabilities that could allow agents to conceal actions or bypass restrictions, prompting the company to say it had fixed the identified weaknesses. The evaluators, METR and Redwood Research, said they had not independently verified the fixes. Two researchers, one from each organization, spent three weeks at OpenAI from July 21 to August 11, 2026. They reviewed internal code and documents and interviewed about a dozen employees as part of a pilot audit of the company’s monitoring system for agent behavior. The assessment covered the system as it existed on and before July 21, and did not evaluate changes made afterward. The researchers found methods that could hide hundreds of an agent’s actions from monitoring and ways to bypass Codex sandboxing, a digital boundary intended to limit what an agent can access. They also identified weaknesses in monitoring triggers and found that specially crafted inputs could likely crash the monitor, according to METR’s account. The audit examined whether monitoring covered agent activity and whether agents could disable or subvert oversight. METR...