On October 22, AI safety evaluation non-profit METR published findings from a red-teaming exercise assessing OpenAI’s internal monitoring system for autonomous AI agents (metr.org). The evaluation was conducted in collaboration with Redwood Research between July 21 and August 11, 2026, inspecting internal safety architecture as it existed in July rather than current production systems.
What the Researchers Evaluated
The evaluation examined internal agent workloads routed through the Responses API for models at the capability level of GPT-5.5 and above. The researchers estimated that 75% to 94% of internal agent traffic was suitable for automated monitoring, while for a specific internal Codex agent deployment, coverage eligibility reached 99.2%. The authors emphasized that these figures represent coverage criteria rather than the fraction of traffic actively inspected.
Broader debates on AI capabilities and safety pacing were explored in our analysis: Why OpenAI and Anthropic Propose Slowing Down AI — And Whether the Race Can Be Paused.
Where the Monitoring Gap Emerged
The exercise identified that due to context compression events not being logged properly, approximately 10% of eligible internal Codex agent traffic between May and July may have operated without monitoring. The precise share remains uncertain due to gaps in historical records. OpenAI had identified this logging issue prior to the evaluation and deployed a fix while the exercise was underway. The report confirms no evidence of user data exfiltration or active exploits.
Innovations in developer tools and local weights were discussed in our report: Recursive Releases Flint for Local Coding with Weights Available under Apache 2.0.
What This Means for Readers
- The summer exercise reflects historical internal infrastructure and does not evaluate or indicate vulnerabilities in current consumer ChatGPT or commercial API services.
- The report highlights the technical challenges in achieving end-to-end oversight across multi-turn autonomous agent environments.
Measures Announced by OpenAI
In a companion public statement (openai.com), OpenAI outlined remedial actions including tighter workload isolation, expanded execution safeguards, and continuous monitoring pipelines. The company noted mandatory chain-of-thought (CoT) monitoring requirements for reinforcement learning training and tool evaluations on models class GPT-5.6 Sol and above.
