25% of OpenAI's GPT-5.5 Traffic Went Unchecked

An external audit of OpenAI has revealed significant gaps in its internal monitoring system. The audit, conducted by researchers from METR and Redwood Research, found that between 6 and 25 percent of traffic from models in the GPT-5.5 class and above was not monitored at all during the audit period. The researchers were given access to OpenAI's codebase and internal documents and spoke with about a dozen employees over three weeks. OpenAI has since fixed all of the weaknesses identified in the audit, but the full report was redacted before being shared with the public.

The audit also found that internal Codex traffic was not fully monitored due to issues with compaction events. A separate gap affected this traffic, as the system did not handle compaction events correctly, leaving about 10 percent of it untracked. OpenAI found the bug independently and fixed it while the audit was still under way.

The auditors also identified further problems, including ways to hide hundreds of actions from the monitor and to bypass the Codex sandbox. They concluded that specially crafted inputs could probably crash the monitor entirely, and that model behavior during reinforcement learning (RL) training was not monitored at all when the audit began. OpenAI paused RL training after the Hugging Face incident in August and expanded monitoring to include it, but these changes came after the auditors' access period had ended.

OpenAI says it has fixed all of the weaknesses the auditors found. The complete report ran to 61 pages, but after redactions, only a 38-page version was shared with METR and Redwood Research. A short summary was released to the public. METR said it had no right to disclose the redacted portions on its own.

For future audits, METR says it wants to push for greater public transparency.