OpenAI Audit Finds Up to 25 Percent of Advanced AI Agent Traffic Unmonitored
External auditors from METR and Redwood Research spent three weeks on-site at OpenAI and found significant gaps in the company's monitoring of AI agents, with up to a quarter of traffic from its most advanced models going unchecked. Large parts of the 61-page report were redacted by OpenAI.
An on-site audit of OpenAI's internal safety monitoring has found that as much as 25 percent of traffic from the company's most advanced AI agents goes unmonitored. The review was carried out by external auditors from METR and Redwood Research, who spent three weeks inside the company examining how it tracks the behaviour of AI agents.
The findings point to a substantial blind spot in the oversight of systems at the GPT-5.5 class and above, the most capable models covered by the audit. The auditors described the gaps as significant, suggesting that a meaningful share of agent activity at the frontier of the company's technology is not being captured by its internal monitoring systems.
Only a summary of the audit has been made public. OpenAI redacted large portions of the 61-page report, leaving the full scope of the auditors' observations and recommendations outside public view. The decision to withhold most of the document limits independent scrutiny of both the problems identified and the company's response to them.
The audit was conducted by two organisations with established roles in AI safety research. METR, formerly the Alignment Research Center, evaluates frontier AI systems and their risks, while Redwood Research focuses on technical work aimed at ensuring advanced AI systems behave as intended. Their presence on-site at OpenAI over three weeks gave them direct access to the company's internal monitoring arrangements for AI agents, software that can carry out tasks with limited human involvement.
Monitoring is a central line of defence as AI agents take on more autonomous work. If a quarter of traffic from the most capable models escapes oversight, unusual or harmful behaviour may go unnoticed, and the record needed to investigate incidents after the fact may be incomplete. The audit's findings therefore raise questions about how far internal safeguards keep pace with the deployment of increasingly powerful systems.
The limited disclosure adds a further dimension. A 61-page report reduced to a public summary leaves researchers, regulators and the wider public unable to assess the auditors' methodology or the severity of the gaps they found. For a company whose systems are used across a growing range of tasks, the balance between commercial confidentiality and transparency about safety failures is likely to remain a point of contention.
The results also feed into a broader debate about how AI safety claims are verified. External audits are increasingly proposed as a way to give assurances that go beyond companies' own statements, but their value depends on access and on how much of the findings are published. An audit that confirms serious gaps while keeping most of its evidence private offers only a partial picture.
OpenAI has not publicly detailed the specific causes of the monitoring shortfall or what steps it intends to take in response. The summary confirms the scale of the unmonitored traffic and the redactions, but leaves open how the gaps arose and whether they have since been closed. With advanced agent systems continuing to be deployed, the audit places the effectiveness of internal oversight at the centre of the discussion about responsible development.
4
