METR red-teams OpenAI's internal agent monitoring and finds coverage gaps
AIMETR and Redwood Research ran a three-week pilot red-teaming exercise on OpenAI's internal misalignment monitoring system in July and August. They estimated that roughly 75–94% of eligible agentic Responses API traffic from models at least as strong as GPT-5.5 was eligible for monitoring, and found vulnerabilities including monitoring gaps from compaction events and a sandbox bypass.
AIWhy it matters.
The post reports concrete monitoring gaps and fixes found in OpenAI's agent oversight, clarifying what external red-teaming can and cannot establish about internal safeguards.
Loading source posts…
Source details
METR Blog · metr.orgPublished · added here
Original title: Red-Teaming OpenAI's Internal Agent Monitoring Systems