Audit: 25% of OpenAI's top model traffic escaped internal checks

OpenAI's internal safety monitoring systems left a portion of its most advanced AI traffic unchecked, according to an external audit released today. Researchers from the evaluation group METR and the research organization Redwood Research conducted the review to assess how effectively the company tracks model behavior.

The audit team was granted deep access to OpenAI's codebase and internal documents between July 21 and August 11. During this three-week on-site evaluation, the auditors spoke with approximately a dozen company employees. The exercise served as a pilot program designed to allow outside researchers to inspect internal infrastructure directly.

The central finding of the report concerns the volume of unmonitored traffic. The researchers estimated that between 6 and 25 percent of internal agentic Responses API traffic from GPT-5.5 class models and above was not monitored at all. METR noted that the uncertainty surrounding this specific range is considerable in both directions.

A separate gap affected internal Codex traffic that was intended to be under surveillance. METR estimates that roughly 10 percent of this traffic went untracked between May and July due to a bug involving compaction events, which are context summaries. OpenAI identified and fixed this bug while the audit was still underway.

The auditors also identified structural weaknesses in how monitoring is triggered. The system only activated when an agent handed control back to the user, allowing agents that continued executing actions to bypass oversight. The researchers demonstrated methods to hide hundreds of actions from the monitor and to bypass the Codex sandbox. They also concluded that specially crafted inputs could likely crash the monitor entirely.

Model behavior during reinforcement learning training was not monitored at the start of the audit. OpenAI stated it paused this training after an incident in August and expanded monitoring to include it, though these changes occurred after the auditors' access period ended.

OpenAI says it has fixed all weaknesses identified in the audit. The full 61-page report was shared internally, while METR and Redwood Research leadership received a redacted 38-page version. A short public summary was released, with METR stating it had no right to disclose the redacted portions independently.