OpenAI's Safety Monitoring Found Up to 25% Unsupervised

An investigation by external audit organizations METR and Redwood Research has revealed critical deficiencies in OpenAI's AI agent monitoring system. After spending three weeks at OpenAI's offices examining internal systems, both organizations confirmed that up to 25% of traffic from agents using GPT-5.5 class and higher models remains outside the scope of monitoring.

An AI agent refers to an AI system capable of autonomously performing tasks when given a goal, without requiring step-by-step human instruction. Because these agents independently determine and execute a wide range of actions—including web searches, code execution, and manipulation of external tools—their behavior monitoring and management are particularly crucial from a safety perspective. OpenAI's GPT-5.5 class and higher models are considered among the company's most high-performing offerings, and their use in increasingly complex tasks continues to expand.

The audit results were compiled into a 61-page report, but OpenAI has kept the majority of it non-public. Currently, only a summary is available to the general public, and the detailed contents of the report remain inaccessible. Both METR and Redwood Research, which conducted the audit, are external organizations specializing in AI safety assessment.

Regarding AI safety, how accurately a developer can understand and control the behavior of its own models is considered a matter directly linked to industry-wide credibility. Particularly for agent-type AI, which operates autonomously, the risk posed by unexpected behaviors is substantial, and the existence of traffic that escapes monitoring scrutiny raises questions about the actual state of safety management practices. OpenAI's ability to detect and manage such gaps becomes critical to maintaining trust.

The non-public status of the majority of the report also merits attention as a separate concern. External audits derive their value from presenting findings to the public from an independent perspective, but restricted disclosure inevitably raises doubts about the audit's effectiveness. As long as only a summary remains publicly available, society's ability to evaluate what specific findings were identified and what improvements were recommended remains severely limited.

Future attention will likely focus on how OpenAI improves its monitoring systems in response to these findings and the degree of transparency with which it discloses such improvements. As AI agent deployment in society accelerates, the industry faces critical questions about how development companies' safety monitoring systems should function and how to ensure external audits remain effective. These issues have become central to maintaining industry-wide trust.