Audit Finds 6 to 25 Percent of GPT-5.5 Traffic Lacked OpenAI Monitoring

Outside auditors found that between 6 and 25 percent of internal Responses API traffic from GPT-5.5 class models and above was not monitored during a three-week review of OpenAI's oversight systems. METR and Redwood Research researchers carried out the work at OpenAI's offices from July 21 to August 11, with access to code and internal records plus conversations with roughly a dozen employees.

The review also examined internal Codex traffic, where METR estimates about 10 percent went untracked between May and July. The gap traced to mishandling of compaction events, which condense context into summaries. OpenAI identified that bug on its own and repaired it before the audit concluded.

Researchers documented several other routes around oversight. Actions could be concealed from the monitor in groups numbering in the hundreds, and the Codex sandbox could be bypassed. Agents could also avoid a monitor that fired only when control returned to the user, allowing an agent that kept running to skip the check. During reinforcement learning training, model behavior was unmonitored when the audit opened. OpenAI paused that training after an August incident on Hugging Face and later extended monitoring to cover it, steps taken after the access window closed.

OpenAI reports it has repaired every weakness the auditors located. The full report runs 61 pages and was delivered internally. OpenAI applied redactions, after which METR and Redwood Research leadership received 38 pages and the public received only a short summary. METR said it lacks authority to publish the withheld sections on its own.

METR said it will push for wider disclosure in any audit that follows. The work formed a pilot program run jointly by METR, Redwood Research and OpenAI to grant outside researchers access to internal infrastructure.