What the audit found
Independent evaluators METR and Redwood Research spent three weeks on-site at OpenAI examining how the company monitors its AI agents. The result: significant gaps in the monitoring system.
The most striking finding is that up to 25 percent of agent traffic from models at the GPT-5.5 class and above goes unmonitored. In other words, a notable share of what the company's most capable models do as agents falls outside oversight.
Most of the report is hidden
The audit produced a 61-page report. OpenAI redacted large portions of it, and only a summary is publicly available. That makes it hard for outsiders to fully assess the details of the findings and how serious the gaps are.
Why it matters
Agents no longer just chat; they write code, access files and act on systems. Monitoring gaps can mean unwanted agent behavior goes unnoticed. The transparency of independent audits is also critical to judging how solid safety claims really are.
