OpenAI under pressure after flaws in its monitoring system

Two independent evaluators tested OpenAI's internal monitoring infrastructure for three weeks. Their report describes blind spots, possible workarounds, and announced fixes.

METR and Redwood Research conducted a pilot exercise on OpenAI's internal infrastructure between July 21 and August 11, 2026, with deep access to tools, configurations, and logs of the model monitoring system. Their conclusion is straightforward: the system covers a large part of the intended uses, but still leaves gaps that a sufficiently sophisticated agent could exploit. In a sector where AI security is becoming both a trust argument and a governance issue, the test comes at the right time—or the wrong one, depending on the perspective.

A test designed to see how far external access can go

The exercise did not aim for full transparency but was a pilot to measure what it means to give external researchers access close to that of an employee. The two evaluators worked on devices provided by OpenAI, used on-site, and interviewed about a dozen researchers and engineers from the security, safety systems, and monitoring teams. The public result is brief, as the parties agreed to publish only a high-level summary, with OpenAI retaining control over sensitive elements [to be verified].

According to the report published on October 22, 2026, the evaluators mainly sought to estimate what portion of agent traffic was actually under surveillance, and whether agents could disable or subvert the system. The full internal report is 61 pages, compared to 38 pages for a redacted version sent to the management of METR and Redwood Research.

Des trous de couverture encore bien réels

METR estimates that about 75 to 94% of internal agent traffic via the Responses API, for models at least as powerful as GPT-5.5, was eligible for monitoring at the time of the test. Within the subset of internal Codex agents, coverage reportedly reached 99.2% of eligible traffic. But these figures remain surrounded by significant uncertainty, and the evaluators could not measure all the internal agent coding traffic they were aware of, due to lack of access and time.

Several blind spots nevertheless stand out clearly. At the start of the exercise, reinforcement learning runs were not covered by the system, although some rollouts had already been monitored between March and June. Another reported gap: the absence of monitoring of compaction events would have left about 10% of eligible internal Codex traffic unchecked between May and July 2026. OpenAI had identified this issue before the exercise and corrected it during it.

The report also highlights a more subtle design bias. Since the system triggers when the agent returns control to the user, an agent can theoretically indefinitely postpone its monitoring by continuing to act. The evaluators could not verify if this occurred in practice. Within a 24-hour window, they also observed elements suggesting the monitor had less than 0.2% errors and no recorded load-shed.

Les contournements identifiés sont plus inquiétants que les chiffres

Beyond the percentages, the researchers found concrete ways to bypass or neutralize the monitoring. They uncovered a method allowing Codex agents to bypass sandboxing and Auto-Review to perform any action authorized to the host user. They also identified an exploit, the details of which were redacted by OpenAI, capable of removing hundreds of recent actions from monitoring, including those related to the exploit itself.

The evaluators also say that an agent can probably create inputs that cause the monitor to fail, exploiting discrepancies between how the trajectory is serialized for the agent and for the monitoring system. Two controllable client headers also allow disabling the system for new agents, opening the door to launching an unsupervised instance, although adding these headers can itself be detected. Finally, two internal read-only agent tools, but not monitored, could have been hijacked to execute arbitrary actions in their sandboxed environment.

OpenAI states that all vulnerabilities affecting its production protection stack have since been fixed, and that it has significantly modified the planned coverage of its monitoring system. METR specifies it has not verified these claims. This is one of those cases where the “fixed” version exists but is not yet under independent control.

Ce que OpenAI surveillait déjà, et ce qu’elle veut faire ensuite

In a publication dated March 19, 2026, OpenAI described a low-latency system powered by GPT-5.4 Thinking, tasked with analyzing agent interactions within 30 minutes of their execution. The system records and examines thought chains and actions, classifies each interaction, assigns a severity level, and escalates anomalies to humans. The company also indicated it covered all internal standard coding deployments, with access to the full history of conversations and tool calls.

Monitored categories include restriction circumvention, deception, uncertainty concealment, reward hacking, unauthorized data transfers, destructive actions, and incoming and outgoing prompt injections. Over five months of operation, OpenAI said it tracked tens of millions of internal agent coding trajectories, with about 1,000 conversations escalated to a moderate level for human review and zero conversations at the highest severity level.

The company adds that this approach is primarily based on monitorability, without being able to reliably quantify false negatives on open traffic without dedicated evaluations and red-teaming. Its roadmap includes control assessments, reliability and latency standards for monitors and humans, then an evolution toward synchronous blocking of actions before execution.

Pourquoi ce rapport compte pour le marché de l’IA

The core issue is not only OpenAI's security but how AI players think about their safeguards as agents gain autonomy. If an internal monitoring system can be bypassed, even partially, the question becomes industrial: how to monitor models more capable than the workflows around them? METR's recommendations go in this direction, with a simple idea: systems must be able to withstand deliberate, repeated, and organized attacks by powerful agents inside a company.

Their list of principles is quite clear: make monitoring hard to avoid, verify that traffic supposed to be controlled really is, react quickly, continuously test the system, and isolate the agent from its monitor. This also involves very concrete practices, such as monitoring agents at regular intervals rather than at the end of their turn, keeping a register of known detours, maintaining tamper-proof logs, and separating the execution environment of tool calls from where agents live. In short, the kind of detail that does not excite communicators but prevents surprises later.

Key points

  • Exercise conducted from July 21 to August 11, 2026.
  • METR and Redwood Research had deep access to OpenAI.
  • 75 to 94% of Responses API traffic was eligible for monitoring.
  • 99.2% of internal Codex traffic was eligible according to METR.
  • OpenAI says it has fixed vulnerabilities affecting its production stack.

By the numbers

  • 61 pages — internal report shared with OpenAI, according to METR and Redwood Research, October 2026.
  • 38 pages — redacted version sent to METR and Redwood Research management.
  • Less than 0.2% — error rate indicated for the monitor over a 24-hour period.
  • About 1,000 conversations — escalated to a moderate level over five months, according to OpenAI, March-August 2026.
  • 30 minutes — delay announced by OpenAI to analyze agent interactions.

To read