OpenAI Faces Scrutiny After Three-Week Audit Exposes Monitoring Blinds Spots

OpenAI is dealing with a new wave of concern after an independent audit revealed serious holes in its safety net. For three weeks, outside researchers had direct access to the company’s internal systems. They found that a large portion of the company’s own AI traffic was not being watched at all. This creates a potential risk where AI agents could act without anyone noticing.

The tension lies in the scale of the oversight gap. While OpenAI operates some of the most powerful AI models in the world, the audit suggests that its internal monitoring was not as comprehensive as assumed. The findings challenge the idea that these systems are under constant, tight control. For a company at the center of the AI race, this is a significant stumble in its safety narrative.

To understand how this happened, we must look at the timeline of the investigation. The audit began on July 21, 2026. Two researchers, one from METR and one from Redwood Research, spent the next three weeks on-site. They worked until August 11. During this time, they were given access to OpenAI’s codebase and internal documents. They also interviewed about a dozen employees to get a full picture of how the systems worked.

The most striking finding concerns the volume of unmonitored traffic. The auditors estimated that between 6 and 25 percent of internal traffic for high-level models, specifically those at the GPT-5.5 class and above, was not covered by monitoring at all. METR noted that this estimate has significant uncertainty, meaning the actual number could be higher or lower. However, even the lower end of that range represents a substantial blind spot in the company’s operations.

There were also specific technical failures in the systems that were supposed to be watching. For internal Codex traffic, the auditors found that roughly 10 percent of activity went untracked between May and July. This was due to a bug involving how the system handled compaction events, which are essentially summaries of context. OpenAI discovered and fixed this bug while the audit was still ongoing. However, during that gap, agents could delay oversight by simply continuing to execute actions instead of handing control back to the user.

The researchers also identified other ways to slip through the net. They found methods to hide hundreds of actions from the monitor. They also discovered ways to bypass the Codex sandbox, a security feature designed to keep AI behavior contained. Furthermore, they determined that specially crafted inputs could likely crash the monitoring system entirely. At the start of the audit, model behavior during reinforcement learning training was not being monitored at all.

OpenAI has responded to these findings. The company says it has fixed all the weaknesses identified by the auditors. It also noted that it paused reinforcement learning training after an incident in August involving Hugging Face. This pause and the subsequent expansion of monitoring happened after the auditors’ access period ended. Therefore, these changes were not part of the official evaluation, but they show the company is acting on the feedback.

Despite the fixes, questions remain about transparency. The full audit report is 61 pages long. OpenAI applied redactions to it. METR and Redwood Research leadership received a 38-page version. The public, however, only received a short summary. METR stated that it did not have the right to disclose the redacted portions on its own. This pilot program was designed to give outside researchers deep access, but the limited public view suggests a balance between openness and internal confidentiality.

What happens next is clear. The audit was a pilot, but it sets a precedent. METR has said it wants to push for more public transparency in future audits. For now, OpenAI has the report and the fixes. The researchers have their findings. The public has a summary. The next step will be seeing if these changes hold up over time. For a company building the future of AI, the margin for error is shrinking.