Independent Audit Exposes Critical Blind Spots in OpenAI’s Internal Safety Controls

The rapid advancement of large language models has intensified scrutiny over how effectively companies monitor their own systems for unexpected behavior. A recent independent audit has highlighted substantial deficiencies in OpenAI’s internal safety infrastructure, raising questions about the current state of oversight for its most advanced models.

The review was conducted from July 21 to August 11, 2026, by two researchers from METR and Redwood Research who were granted on-site access to OpenAI’s facilities. During this three-week period, the auditors examined the company’s codebase and internal documentation while interviewing approximately a dozen employees. The initiative was designed as a pilot program to provide outside experts with deep visibility into the infrastructure that safeguards against rogue AI behavior.

The findings revealed that a significant portion of internal traffic involving the GPT-5.5 class of models and above was not subject to monitoring. METR estimates that between six and 25 percent of this agentic traffic lacked oversight, though the organization noted that the uncertainty surrounding this figure is substantial. Additionally, roughly 10 percent of internal Codex traffic that was intended to be tracked went unmonitored between May and July due to a technical gap in how the system handled compaction events, which are context summaries. OpenAI identified and corrected this specific bug while the audit was still in progress.

Beyond these coverage gaps, the auditors discovered methods to conceal hundreds of actions from the monitoring systems and to bypass the Codex sandbox. They also determined that specially crafted inputs could likely crash the monitor entirely. A further weakness was identified in the handling of reinforcement learning (RL) training, where model behavior was not monitored at the start of the audit. OpenAI stated that it paused RL training following an incident in August and subsequently expanded monitoring, but these adjustments occurred after the auditors’ access period ended.

The distribution of the audit results has also drawn attention. While OpenAI claims to have fixed all identified weaknesses, the full 61-page report was shared internally. After OpenAI applied redactions, the external research leaders received a 38-page version, with only a brief summary released to the public. METR indicated that it did not have the authority to disclose the redacted sections independently.

As OpenAI moves forward, the pilot program aims to refine how external researchers interact with internal infrastructure. For future audits, METR has expressed a desire to advocate for increased public transparency. The company continues to integrate the fixes identified by the auditors, seeking to close the gaps that allowed for potential oversight delays and hidden actions during the review period.

This event marks a significant step in the evolving landscape of AI safety verification. By allowing external experts to probe internal systems, OpenAI has set a precedent for rigorous independent review. The outcome of these efforts will likely influence how other leading organizations approach the monitoring of their most powerful models in the coming months.