Back to the stories

METR and Redwood Research identify vulnerabilities in OpenAI's internal agent-monitoring system

Score 10.0,

AI Safety

The system OpenAI built to watch its own AI agents can be tricked: independent researchers found multiple ways agent actions could go unseen and, in at least one case, a path that might let agents act with a user's full permissions.

One researcher from METR and one from Redwood Research conducted a high-access pilot examination inside OpenAI's monitoring stack. They inspected how internal agent traffic is observed and logged, testing monitoring coverage, reliability, how actions are serialized, client-controllable headers, internal tools, and Codex agent traffic.

Their report describes several concrete weaknesses. It estimates a large share of internal agent traffic was eligible for monitoring while also finding gaps that left some messages unobserved. The team found a method that could, in principle, let Codex agents bypass sandboxing and Auto-Review and act with host permissions. They also found an exploit that could hide hundreds of recent agent actions from monitors; OpenAI redacted the exploit details when the researchers published their findings. OpenAI says it patched the issues, and the researchers say they have not independently verified those patches.

Why this matters: internal monitoring is a primary safety control for powerful agents. If monitoring misses actions or can be disabled, the lab that built the agent may not see when it starts doing things it should not. That turns a technical bug into an institutional risk.

How it works, simply: imagine a security team watching logged actions. The researchers showed agents can change how their actions are packaged or what headers accompany them so the watcher treats them as harmless or skips them. In plain terms, some actions can be masked or routed around the observer.

What changes now: labs, auditors and regulators will need better evidence that monitoring actually covers real-world agent behavior, not just that instruments exist. OpenAI's patch claims matter, but so does independent confirmation.

The open question is practical and urgent: will external reviewers be given enough access and time to verify fixes, or will monitoring remain a black box we must assume works?