OpenAI Audit: 25 Percent of Top Model Traffic Went Unchecked

Up to 25 percent of internal traffic for OpenAI's most advanced AI models bypassed the company's safety monitors, according to a new audit by outside researchers. The estimate, which covers models in the GPT-5.5 class and above, suggests that a large portion of agentic Responses API activity operated without oversight during the review period.

Two researchers from the evaluation group METR and the research organization Redwood Research conducted the on-site audit from July 21 to August 11. They were granted access to OpenAI's codebase and internal documents and interviewed approximately 12 employees. The exercise served as a pilot program designed to provide external parties with deep visibility into the company's internal infrastructure.

The auditors identified a specific technical flaw affecting internal Codex traffic. METR estimates that roughly 10 percent of this traffic went untracked between May and July because the monitoring system failed to process compaction events, which are context summaries. OpenAI independently discovered and fixed this bug while the audit was still in progress. A separate structural issue allowed agents to avoid monitoring entirely by continuing to execute actions without returning control to the user.

The researchers also found that agents could hide hundreds of individual actions from the monitor and bypass the Codex sandbox. They concluded that specially crafted inputs could likely crash the monitor entirely. Additionally, model behavior during reinforcement learning training was not monitored at all when the audit began. OpenAI stated that it paused reinforcement learning training following an incident with Hugging Face in August and subsequently expanded its monitoring coverage.

OpenAI said it has corrected all weaknesses identified by the auditors. The full 61-page report was shared internally, while METR and Redwood Research leadership received a 38-page version with redactions applied by OpenAI. Only a short summary has been released to the public. METR stated it did not have the authority to disclose the redacted sections on its own.

The audit period ended before OpenAI implemented changes to monitor reinforcement learning training, so those updates were not part of the final evaluation. METR indicated it wants to push for greater public transparency in future audit cycles.