OpenAI Faces Scrutiny Over Significant Gaps in AI Safety Protocols

A recent independent audit has exposed significant vulnerabilities in OpenAI’s internal systems for monitoring advanced artificial intelligence, raising questions about the robustness of current safety protocols. The assessment, conducted by external researchers, revealed that a substantial portion of the company’s most advanced AI traffic was operating without adequate oversight, highlighting a tension between rapid development and rigorous safety verification. This discovery comes at a critical juncture for the industry, where trust in autonomous AI systems is increasingly under pressure from both technical experts and the general public.

The audit was carried out from July 21 to August 11, 2026, by two researchers from METR and Redwood Research. During this three-week period, the team was granted deep access to OpenAI’s internal infrastructure, including its codebase and proprietary documents. They also conducted interviews with approximately a dozen company employees to understand the operational context of the monitoring systems. This pilot program was organized jointly by the two research firms and OpenAI to provide an unprecedented level of external visibility into how the company manages its internal AI agents.

The primary focus of the investigation was the monitoring of 'rogue' behavior in agentic systems, specifically those utilizing the Responses API. The auditors determined that an estimated six to 25 percent of internal traffic from models at the GPT-5.5 class and above was not covered by any monitoring during the audit window. METR noted that the uncertainty surrounding this figure is significant in both directions, meaning the true proportion of unmonitored traffic could be higher or lower than the central estimate. This gap suggests that a meaningful segment of the company’s most powerful internal tools was operating in a state of reduced observability.

Further weaknesses were identified in the monitoring of internal Codex traffic. METR estimates that roughly 10 percent of this traffic went untracked between May and July due to a specific technical flaw related to 'compaction events,' which are context summaries used to manage memory. The auditors found that during this period, the monitoring system only triggered when an agent handed control back to a user. Consequently, agents could effectively delay oversight by continuing to execute actions without returning control, thereby sidesteeping the monitoring triggers. OpenAI confirmed that it had independently discovered this bug and implemented a fix while the audit was still underway.

The researchers also uncovered more direct methods for evading oversight. They demonstrated that it was possible to hide hundreds of actions from the monitor and to bypass the Codex sandbox, which is designed to isolate AI operations. Additionally, the team determined that specially crafted inputs could likely cause the monitor to crash entirely. At the start of the audit, model behavior during Reinforcement Learning (RL) training was not being monitored at all. OpenAI stated that it paused RL training following the Hugging Face incident in August and subsequently expanded monitoring to cover this area, although these changes were implemented after the auditors’ access period concluded and were therefore not part of the formal evaluation.

In response to the findings, OpenAI asserts that it has fixed all the weaknesses identified by the auditors. The full report, which spans 61 pages, was handed over to the company internally. However, the distribution of the findings has raised questions about transparency. After OpenAI applied redactions, METR and Redwood Research leadership received a 38-page version of the report. Only a short summary was made available to the general public. METR stated that it did not have the right to disclose the redacted portions on its own, a constraint that has led to calls for greater clarity in future assessments.

This audit serves as a pilot for a new model of external verification in the AI sector. The collaboration between METR, Redwood Research, and OpenAI was designed to test the feasibility of giving outside researchers deep access to internal infrastructure. While the initial results have highlighted areas for improvement, the process itself is seen as a step toward more rigorous industry standards. The findings underscore the complexity of monitoring autonomous systems that can act quickly and independently within complex software environments.

Looking ahead, METR has indicated that it intends to push for increased public transparency in future audits. The organization believes that broader access to detailed findings will help build confidence in the safety of advanced AI systems. As the industry continues to scale, the balance between proprietary secrecy and public accountability remains a central challenge. The next steps will likely involve further refinements to monitoring protocols and potentially new rounds of independent reviews to verify the long-term effectiveness of the fixes implemented by OpenAI.

The implications of these findings extend beyond a single company, touching on broader debates about how the industry manages risk as AI capabilities grow. The identification of specific gaps, such as the handling of compaction events and the lack of monitoring during RL training, provides concrete examples of where safety nets may be insufficient. As OpenAI continues to integrate these fixes, the focus will shift to verifying that these changes hold under real-world conditions, ensuring that the monitoring systems are as robust as the models they are meant to oversee.