External researchers found gaps in internal agent monitoring
METR and Redwood Research spent three weeks testing OpenAI’s internal agent monitoring. They found weaknesses in monitoring coverage and in the conditions that triggered monitoring.
PRACTICAL IMPLICATIONS
What does this mean?
Monitoring systems need independent testing too. Long-running agents and conversation compaction can create gaps when monitoring only starts at particular points.
LIMITS OF THE EVIDENCE
What does this not establish?
The findings concern the system as it existed on 21 July 2026. They do not establish current coverage. OpenAI had already identified some issues and fixed them during the exercise.
Significance: significant
Independent testing provides practical evidence of the limits of monitoring implementations.
An editorial judgement, not a numerical threat forecast.
ORIGINAL PUBLICATION
Red-Teaming OpenAI's Internal Agent Monitoring Systems
METR · published 22 Oct 2026
Read the original source ↗Source checked 13 Nov 2026. Summary updated 13 Nov 2026. Reference date for the period discussed: 21 Jul 2026. This summary was written with AI assistance using the original source.
← Back to the archive