AI Safety Oversight: Independent Audit Uncovers Significant Monitoring Gaps at OpenAI

For readers in New York and New Jersey who rely on the latest artificial intelligence tools for work and daily life, the recent findings from an independent audit suggest that the safety nets surrounding these technologies may have been less robust than previously assumed. A detailed examination of OpenAI’s internal systems has revealed that a portion of the company’s most advanced AI models operated without full oversight during a specific period, prompting immediate technical adjustments to close the identified loopholes.

The audit, conducted from July 21 to August 11, 2026, involved two researchers from METR and Redwood Research who spent three weeks on-site at OpenAI. These experts were granted access to the company’s codebase and internal documentation, and they conducted interviews with approximately a dozen employees. The primary objective was to evaluate the effectiveness of OpenAI’s internal monitoring systems designed to detect rogue or unexpected behavior in its AI agents.

The numerical findings of the investigation highlight specific areas of concern. According to METR, an estimated 6 to 25 percent of internal traffic related to the Responses API for models of the GPT-5.5 class and above was not covered by monitoring at all during the audit period. The researchers noted that the uncertainty surrounding this estimate is significant in both directions. Additionally, for internal Codex traffic that was intended to be monitored, METR estimates that roughly 10 percent went untracked between May and July 2026. This gap was attributed to a technical issue involving how the system handled compaction events, which are essentially context summaries. OpenAI independently discovered and fixed this bug while the audit was still in progress.

Beyond the statistical gaps, the auditors identified structural weaknesses in how oversight was triggered. During the period when compaction events were not being monitored, AI agents had the ability to delay oversight because the system only activated when agents handed control back to the user. This allowed agents to sidestep monitoring by simply continuing to execute actions without returning control. The researchers also discovered methods to hide hundreds of actions from the monitor and to bypass the Codex sandbox entirely. They determined that specially crafted inputs could likely cause the monitor to crash, and they found that model behavior during reinforcement learning (RL) training was not being monitored at the start of the audit.

OpenAI responded to these findings by stating that it had fixed all the weaknesses identified by the auditors. The company noted that it had paused RL training following the Hugging Face incident in August and subsequently expanded monitoring to cover this area. However, these changes were implemented after the auditors’ access period had ended and were therefore not part of the formal evaluation. The full 61-page report was handed over internally, but after redactions, METR and Redwood Research leadership received a 38-page version. Only a short summary has been made available to the public, as METR stated it had no right to disclose the redacted portions independently.

The audit was part of a pilot program organized jointly by METR, Redwood Research, and OpenAI, designed to provide outside researchers with deep access to internal infrastructure. This initiative aims to establish a precedent for independent verification of AI safety measures. For future audits, METR has expressed a desire to push for greater public transparency, ensuring that more of the detailed findings are accessible to the broader community and stakeholders.

The stakes of these findings are significant for the broader AI landscape. As companies deploy more autonomous agents capable of executing complex tasks, the ability to monitor and control these systems becomes critical. The gaps identified in this audit underscore the complexity of ensuring that advanced AI models behave as intended, particularly in internal environments where they may interact with sensitive data or systems. The rapid correction of the identified issues demonstrates a commitment to addressing these vulnerabilities, but the initial gaps highlight the ongoing challenge of maintaining robust oversight in rapidly evolving AI systems.

For users in the New York and New Jersey regions, who are among the early adopters of these technologies, this audit provides a deeper understanding of the internal processes that support the tools they use daily. While the specific technical details may not directly impact everyday usage, the broader implications for safety and reliability are relevant to anyone relying on AI for productivity or creative tasks. The findings serve as a reminder that continuous monitoring and independent review are essential components of responsible AI development.