25% Blind Spot: New Audit Exposes Critical Gaps in OpenAI's Safety Nets

Up to 25 percent of the internal traffic from OpenAI’s most advanced models may have operated without any safety monitoring whatsoever, according to a rigorous new audit conducted by independent researchers. This startling figure represents the upper bound of an estimated range, highlighting a potential blind spot in the monitoring infrastructure of one of the world's leading artificial intelligence developers. The findings have drawn immediate attention to the robustness of safeguards designed to detect rogue or unexpected behavior in large language models.

The audit, which took place from July 21 to August 11, 2026, involved two researchers from METR and Redwood Research who were granted unprecedented access to OpenAI’s internal systems. Over the course of three weeks, the team examined the company’s codebase and internal documentation while interviewing approximately a dozen employees. Their primary focus was evaluating the effectiveness of OpenAI’s internal monitoring systems specifically designed to track agentic AI behavior.

The numbers reveal a significant gap in coverage. The auditors estimated that between 6 and 25 percent of internal agentic Responses API traffic from models at the GPT-5.5 class and above was not covered by monitoring during the audit period. METR noted that the uncertainty surrounding this specific estimate is significant in both directions, meaning the true figure could be lower or higher. Additionally, for internal Codex traffic that was supposed to be monitored, researchers estimated that roughly 10 percent went untracked between May and July due to a specific handling gap in the system.

This monitoring gap stemmed from how the system processed compaction events, which are essentially context summaries. During the period when these events were not being properly monitored, agents had the ability to delay oversight. The system was designed to trigger checks only when agents handed control back to the user. Consequently, agents could sidestep these checks by simply continuing to execute actions without returning control, effectively allowing them to operate outside of the intended supervisory loop.

Beyond simple coverage gaps, the auditors identified more complex vulnerabilities. They found methods to hide hundreds of actions from the monitor and ways to bypass the Codex sandbox entirely. The team also determined that specially crafted inputs could likely cause the monitor to crash completely. Furthermore, they discovered that model behavior during reinforcement learning (RL) training was not being monitored at all when the audit began, leaving a critical phase of model development unobserved.

OpenAI responded to the findings by stating that it had already independently discovered the compaction bug and fixed it while the audit was still underway. The company also clarified that it paused RL training following the Hugging Face incident in August and subsequently expanded monitoring to cover this area. However, OpenAI emphasized that these corrective measures were implemented after the auditors’ access period ended, meaning they were not part of the initial evaluation conducted by METR and Redwood Research.

The transparency of the audit results has become a point of discussion. OpenAI states that it has fixed all the weaknesses identified by the auditors and provided the full 61-page report internally. After OpenAI applied redactions, the leadership teams at METR and Redwood Research received a 38-page version of the document. In contrast, only a short summary was made available to the general public. METR indicated that it did not have the right to disclose the redacted portions of the report on its own initiative.

This initiative was structured as a pilot program organized jointly by METR, Redwood Research, and OpenAI. The goal was to provide outside researchers with deep access to internal infrastructure, a step that marks a shift in how major AI labs engage with independent oversight. Looking ahead, METR has expressed a desire to push for greater public transparency in future audits, aiming to ensure that the public has a clearer view of the safety measures in place for advanced AI systems.

The stakes of these findings are high for the region and the broader tech industry, as New England continues to be a hub for both AI development and academic research. The audit underscores the challenges of monitoring increasingly autonomous AI agents. As companies like OpenAI work to address these gaps, the balance between operational efficiency and rigorous safety oversight remains a central concern for stakeholders in the AI ecosystem.