Independent Audit Finds Monitoring Gaps at OpenAI

As the regulatory landscape for artificial intelligence continues to evolve, questions regarding the robustness of internal safety mechanisms have grown increasingly urgent for industry stakeholders.

A three-week on-site evaluation conducted by researchers from METR and Redwood Research has identified substantial deficiencies in OpenAI’s systems for monitoring the behavior of its advanced models. The independent team was granted comprehensive access to the company’s codebase and internal documentation between July 21 and August 11, 2026. During this period, the auditors also conducted interviews with approximately a dozen OpenAI employees to better understand the operational context of the monitoring infrastructure.

The primary finding of the investigation is that a significant portion of internal traffic was not subject to oversight. The researchers estimate that between six and 25 percent of agentic Responses API traffic involving models of the GPT-5.5 class or higher was entirely unmonitored. METR noted that the uncertainty surrounding this figure is considerable in both directions. Additionally, for internal Codex traffic that was intended to be tracked, the audit revealed that roughly 10 percent went unrecorded between May and July due to a specific gap in how the system processed compaction events, which function as context summaries. OpenAI independently identified and corrected this bug while the audit was still in progress.

Beyond these quantitative gaps, the auditors discovered qualitative vulnerabilities that could allow agents to evade oversight. During periods when compaction events were not monitored, agents could delay scrutiny by continuing to execute actions without returning control to the user, effectively bypassing the trigger for oversight. The team also demonstrated methods to conceal hundreds of actions from the monitor and to circumvent the Codex sandbox. Furthermore, they determined that specially crafted inputs could likely cause the monitor to crash entirely, and they found that model behavior during reinforcement learning (RL) training was not being monitored at the start of the audit.

OpenAI stated that it paused RL training following an incident with Hugging Face in August and subsequently expanded monitoring to include this area. However, the company acknowledged that these changes were implemented after the auditors’ access period concluded, meaning they were not part of the official evaluation. While OpenAI asserts that all identified weaknesses have now been fixed and that the full 61-page report was provided internally, the level of public disclosure remains limited.

The full report underwent redaction by OpenAI before being shared with the leadership of METR and Redwood Research, who received a 38-page version. Only a brief summary has been made available to the general public. METR stated that it did not have the authority to disclose the redacted portions independently. This audit served as a pilot program designed to provide outside researchers with deep access to internal infrastructure, with METR expressing a desire for greater public transparency in future evaluations.