External auditors from METR and Redwood Research conducted a three-week on-site check at OpenAI. They studied the internal monitoring system and reported substantial gaps: up to 25% of streams from GPT-5.5 class and above, according to the description, remained unmonitored.
OpenAI covered significant parts of the 61-page report, while publicly available only its summary. This limits the ability to independently assess the audit methodology, the full scale of identified issues, and the company’s response.
For AI safety, the key concern is not only the existence of monitoring but its actual coverage. In this case the conclusions are based on a single independent summary by The Decoder, not on the published full report.
editorial commentary
Why it matters
Likely consequence — increased attention to the completeness of monitoring and publication of audit materials. The next observable signal will be OpenAI’s response or the disclosure of the full report. Substantial uncertainty remains due to editorial omissions and the absence of the primary document.