Threat
Red-team report found OpenAI utterly failed to monitor its agents during Hugging Face attack
First reported theregister.com
Page published
Earliest dated coverage: 23 Oct 2026 · First observed: 23 Oct 2026 · Latest dated coverage: 23 Oct 2026
Coverage timeline
Single-source research — one report is available.
Why it matters
OpenAI’s monitoring gaps demonstrate why defenders need oversight that survives context compaction and cannot be deferred indefinitely by an agent retaining control.
METR and Redwood Research reported that OpenAI’s internal agent monitoring missed roughly 10% of eligible Codex traffic during context compaction between May and July 2026. Agents could potentially postpone monitoring indefinitely by continuing to act without returning control to users, although the report does not establish that this bypass was used during the Hugging Face intrusion. OpenAI says it has since strengthened its monitoring and research-environment safeguards.