Three-week on-site audit at OpenAI reveals major gaps in AI safety monitoring

A three-week on-site audit conducted by external researchers from METR and Redwood Research at OpenAI uncovered notable shortcomings in the company's internal monitoring infrastructure for AI agents. The review focused on how OpenAI tracks and supervises the behavior of its most capable models, specifically those at the GPT-5.5 class and above.

According to the auditors, as much as a quarter of agent traffic from these frontier-tier models was not being monitored at all. This means that a substantial portion of autonomous agent activity was operating outside the scope of OpenAI's safety oversight systems, raising questions about the organization's ability to detect and respond to risky behavior in real time.

The finding is significant because AI agents represent a growing deployment category where models take multi-step actions with limited human intervention. Effective monitoring is a prerequisite for safe scaling, and gaps of this magnitude could undermine confidence among regulators, enterprise customers, and the broader AI safety community. The involvement of independent external auditors lends additional credibility to the assessment.

The audit highlights a broader tension in the AI industry between rapid capability development and the maturation of safety tooling. As frontier labs push toward more autonomous systems, the infrastructure needed to observe, constrain, and audit those systems must keep pace. This review suggests that at least one major lab may be behind on that front, and it could prompt similar scrutiny at other organizations deploying comparable agent capabilities.

This page provides an editorial summary based on publicly available information. It is not a republished article. Use the source link below for the original report.