External auditors from METR and Redwood Research spent three weeks on site at OpenAI examining how the company monitors its AI agents internally. According to The Decoder, they found significant gaps: up to 25 percent of agent traffic from the GPT-5.5 class onward was presumably not monitored.

What the audit shows

AI agents are systems that do more than answer questions. They carry out steps on their own, such as writing programs, operating tools or working through tasks over longer periods. Because they have a lot of room to act, developers rely on monitoring systems. These are meant to detect conspicuous or risky actions before they cause harm.

If a substantial share of the data traffic never reaches this system in the first place, a blind spot emerges. The value of monitoring therefore depends not only on how well it detects problems, but also on whether it covers all relevant activity. According to the source, the figure of up to 25 percent is based on an assumption by the auditors.

Little transparency

The report is 61 pages long. OpenAI has redacted large parts of it, and only a summary is publicly available. For outsiders, it is therefore hard to tell exactly which gaps existed, how they arose and whether they have since been closed. The source gives no information on this.

More on this topic:

Why this matters

Independent on-site audits are seen as a way to check the safety promises of AI companies instead of relying solely on their self-reporting. This case shows that such audits can bring weaknesses to light. At the same time, their usefulness to the public depends on how much is ultimately published.

Anyone who uses AI agents in a company can take away a question from this: how complete is the monitoring in practice, and can that be verified independently? For details, it is worth reading the published summary yourself and watching for further information.