Audit Exposes Blind Spots in OpenAI's Internal Safety Nets

An estimated 25 percent of internal traffic from OpenAI’s most advanced models may have operated without any safety oversight during a recent three-week audit. This upper-bound figure, derived from a review of the company’s monitoring infrastructure, highlights a critical vulnerability in the systems designed to detect rogue behavior in large language models. The assessment, conducted by researchers from METR and Redwood Research, suggests that the gap in coverage was not merely theoretical but a structural flaw in how OpenAI tracked its own agentic workflows.

The timeline of this discovery begins in late July 2026, when the audit team was granted deep access to OpenAI’s codebase and internal documentation. From July 21 to August 11, the researchers interviewed approximately a dozen employees and examined the mechanisms used to monitor the Responses API. Their focus was specifically on models at the GPT-5.5 class and above, which represent the frontier of the company’s current capabilities. This period coincided with a critical window where the company was refining its internal tools, yet the auditors found that between six and 25 percent of this high-stakes traffic was entirely invisible to the monitoring systems. METR noted that while the central estimate is concerning, the uncertainty surrounding that figure is significant in both directions.

Beyond the broad coverage gaps, the audit uncovered specific technical failures that allowed agents to evade oversight. For internal Codex traffic, which was supposed to be fully monitored, roughly 10 percent of actions went untracked between May and July due to a bug involving compaction events, or context summaries. During this period, agents could delay scrutiny simply by continuing to execute actions rather than handing control back to the user. The auditors also identified methods to hide hundreds of actions from the monitor and bypass the Codex sandbox. Furthermore, they determined that specially crafted inputs could likely crash the monitoring system entirely, and that model behavior during reinforcement learning (RL) training was not being monitored at the start of the evaluation.

OpenAI responded by stating that it had independently discovered the compaction bug and fixed it while the audit was still underway. The company also paused RL training following an incident in August and expanded monitoring to cover that process. However, these corrective measures were implemented after the auditors’ access period ended, meaning they were not part of the final evaluation. Despite these fixes, the lack of visibility during the audit window remains a key finding, illustrating how easily oversight can be circumvented in complex agentic systems.

The transparency of the findings has become a point of contention. While OpenAI asserts that all identified weaknesses have been corrected, the full 61-page report was redacted before being shared with the auditors. METR and Redwood Research leadership received a 38-page version, but only a short summary was released to the public. METR stated it had no contractual right to disclose the redacted portions independently, limiting the public’s ability to verify the extent of the issues.

This audit served as a pilot program for a new model of external verification, where independent researchers gain direct access to a lab’s internal infrastructure. For future iterations of this process, METR is advocating for greater public transparency to ensure that the public can assess the safety measures protecting these powerful systems. The findings underscore the difficulty of monitoring autonomous systems and suggest that current internal safeguards may be insufficient without rigorous, independent scrutiny.

As the industry moves toward more agentic and autonomous AI applications, the results of this audit provide a stark reminder of the challenges involved in ensuring safety at scale. The gap between intended monitoring and actual coverage highlights the need for more robust, real-time oversight mechanisms. As OpenAI and other labs continue to scale their operations, the pressure to close these gaps will only intensify, making the push for transparent auditing a central issue in the ongoing debate over AI governance.