Independent Audit Exposes Blind Spots in OpenAI's Internal Safety Monitoring

For ordinary users interacting with OpenAI’s advanced language models, the recent findings suggest that the safety nets designed to catch unexpected behavior may have been thinner than previously understood. While the company has addressed the specific issues identified, the audit raises questions about how thoroughly internal systems were being watched during a critical period of development in the second half of 2026. The implications are not immediate changes to consumer interfaces, but rather a shift in the internal rigor applied to the models that power those services. The core issue is one of visibility: if the monitoring systems themselves have blind spots, the assurance that rogue behaviors are being caught in real-time is compromised.

The scrutiny came from researchers at METR and Redwood Research, who spent three weeks on-site at OpenAI from late July to mid-August 2026. This access allowed them to examine the codebase, review internal documents, and interview roughly a dozen employees. Their primary focus was the company's internal monitoring infrastructure, specifically designed to detect anomalous or rogue behavior in agentic systems. The audit was structured as a pilot program to test the viability of granting outside experts deep access to such sensitive infrastructure, a move that signals a potential new standard for AI safety verification.

One of the most significant findings involves a substantial gap in coverage for high-level model traffic. The auditors estimated that between 6 and 25 percent of internal traffic using the agentic Responses API from models of the GPT-5.5 class and above was entirely unmonitored during the review period. METR noted that the uncertainty surrounding this figure is significant in both directions, meaning the actual percentage could be higher or lower. This gap suggests that a meaningful slice of the company's most advanced internal operations was proceeding without the intended safety oversight during that specific window.

Beyond the coverage gap, the auditors identified technical vulnerabilities in how the system handled data compaction, which involves summarizing context to manage memory. For internal Codex traffic, roughly 10 percent of sessions between May and July 2026 went untracked due to a bug in how these compaction events were processed. OpenAI had independently discovered and fixed this bug while the audit was still ongoing. Furthermore, the system’s trigger mechanism allowed agents to delay oversight by simply continuing to execute actions rather than handing control back to the user, effectively letting them sidestep checks. The researchers also demonstrated ways to hide hundreds of actions from the monitor and bypass the Codex sandbox, while noting that specially crafted inputs could likely crash the monitoring system entirely.

The audit also highlighted that model behavior during reinforcement learning (RL) training was not monitored at all when the inspection began. OpenAI stated that it paused RL training following an incident involving Hugging Face in August and subsequently expanded monitoring to cover this phase. However, these changes were implemented after the auditors' access period concluded, meaning they were not part of the formal evaluation. The company asserts that it has now fixed all the weaknesses identified by the external team. The full 61-page report was shared internally, but after redactions, the external researchers received a 38-page version, with only a brief summary made public. METR indicated it had no right to disclose the redacted sections independently, though it aims to push for greater transparency in future audits.

This pilot program marks a potential shift in how AI safety is verified, moving from internal self-assessment to external scrutiny. The findings do not imply that the models were unsafe during the unmonitored periods, but rather that the company lacked the tools to know for sure in real-time. As OpenAI continues to scale its operations, the pressure to maintain comprehensive monitoring will likely increase. The next phase of these audits will be critical in determining whether this new level of transparency becomes a permanent feature of the industry's approach to AI safety.