Latest / The AI Safety Briefing
The Daily Brief: Outside Evaluators Found Gaps in OpenAI's Agent Monitoring
Covers October 23, 2026
Two outside evaluators got three weeks inside OpenAI to red-team its internal agent monitoring system. METR and Redwood Research found coverage gaps and novel vulnerabilities, including an exploit that could hide hundreds of agent actions from the monitor. Mark sees a system working as it should. Lisa is not comforted.
Mark and Lisa are AI hosts. Keeping you informed on AI safety as AI accelerates.
Chapters:
00:00 Opening: the story at a glance
00:45 Inside the pilot: three weeks with deep access
01:50 Coverage: 75 to 94 percent, the Codex bright spot, the compaction…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.