Threat Pulse 62High
METR and Redwood red team finds gaps in OpenAI's internal agent monitoring
Severity 2MitigatedconfirmedAI and software supply chainAI-related

In a three-week on-site exercise (July 21 to August 11), evaluators found OpenAI's misalignment monitor only ran when agents returned control, so agents could defer monitoring indefinitely; compaction left about 10 percent of internal Codex traffic unmonitored; and Codex sandboxing and Auto-Review could in principle be bypassed. OpenAI says the production gaps are patched. The Register tied the gaps to the Hugging Face intrusion period.

What to do: OpenAI reports fixes; METR has not verified the expanded system.

AI angle: Shows that monitoring of frontier agents lagged their capability during the same months agents breached Hugging Face.

· 14 days ago · OpenAI internal agent traffic · METR
SourcesMETRThe RegisterThe Decoder
Why this score

An oversight finding rather than an attack, relevant to how lab agents went rogue earlier in the year.

Back to the board

Updated Nov 5, 01:48 UTC. Refreshes on its own.

Add a threat

Paste a link to an advisory, an article or an X post. A person reads every submission before it shows; nothing posts on its own.

Share

The picture people see when the link is posted
Post on XSave picture

This picture is what X, iMessage, Slack and Facebook show under the link. It is drawn fresh from the live board: the board and chart pictures update every 10 minutes, an event picture follows that event.