OpenAI AI Agent Monitoring Gaps Found in Three-Week External Audit

What happened

External auditors from METR and Redwood Research found significant gaps in the internal monitoring system OpenAI uses for its AI agents, with up to 25 percent of agent traffic from models at the GPT-5.5 class and above going unmonitored. According to The Decoder, the two organisations spent three weeks on-site at OpenAI reviewing that system.

The findings are contained in a 61-page report. OpenAI redacted large portions of it, so only a summary is publicly available. That means the detailed methodology and the specific weaknesses identified cannot be independently assessed from the public material alone.

The audit focused on how OpenAI watches what its agents do internally, rather than on the models' outputs in isolation. The on-site timeframe and the external identity of the reviewers are the key facts reported so far.

Why it matters

As organisations hand more work to autonomous AI agents, the question of who observes those agents, and how completely, moves from a technical detail to an operational one. A monitoring system that covers only part of the traffic leaves the remainder outside the visibility that teams rely on to catch errors, misuse or unexpected behaviour. The figure reported here, a quarter of traffic from the most capable model class, shows how quickly coverage can lag behind deployment.

For leaders in AI, service design and digital transformation, the lesson extends beyond one company. Any organisation rolling out agents into customer-facing or internal workflows will need to know what share of agent activity its oversight actually reaches, and whether that share is measured at all. Independent review, and the transparency to support it, is becoming part of how that assurance is established.

By the numbers

• Up to 25 percent of agent traffic from models at the GPT-5.5 class and above reportedly went unmonitored.
• Three weeks was the length of the on-site audit at OpenAI by METR and Redwood Research.
• 61 pages is the length of the full audit report, of which only a redacted summary is public.

The Renascence take

The headline gap is important, but the more revealing detail is how little of the report can be read. Redaction at this scale turns oversight into something customers and regulators must take on trust, and that is a service-design problem as much as a safety one.

Most readers will focus on the 25 percent figure, yet the real signal is that coverage was not visible until outsiders looked. Trust in any system depends on people being able to see what is and is not being watched. Operators should therefore track monitoring coverage as a headline metric, publish it in plain terms, and treat any gap between deployment volume and oversight volume as a product defect rather than a footnote.