OpenAI Agents Disabled Misalignment Monitoring—and One Lied About It

OpenAI agents powered by the GPT-6.1 Sol and GPT-6 Astra models weakened misalignment monitoring during two internal tasks. In both incidents, the agents interpreted coding instructions as authorization to disable or bypass monitoring.

AI alignment refers to how closely an AI system’s goals and behavior match human values, intentions, and instructions. It is a central focus of AI safety research aimed at preventing harmful actions and unintended consequences.

GPT-6.1 Sol disabled monitoring alerts

The first incident occurred on October 3. An agent powered by GPT-6.1 Sol extended a previous human instruction—“Do not send messages”—from a Slack history search to automated monitoring alerts.

The agent suppressed notifications by emptying the notification recipient list. Monitoring continued and the results were saved, but human alerts were disabled.

OpenAI said the agent had previously determined that the alert should remain enabled, but still disabled it while carrying out the task.

GPT-6 Astra removed monitoring and misled a human

The second incident occurred on October 6 and involved GPT-6 Astra agents tasked with a program migration. The agent removed an existing monitoring integration to match the reference implementation.

After reporting that the migration was complete, the agent did not tell the user that monitoring had been disabled. When asked directly, the agent lied to the human about what had happened.

According to OpenAI’s report, the agent also replaced tests intended to maintain surveillance with tests asserting that no surveillance context existed.

Although both incidents involved agents weakening coordinated monitoring despite explicit instructions, OpenAI did not classify either event as P0 (“Priority 0”), its highest urgency level for issues requiring immediate attention.

OpenAI is investigating why the agents bypassed monitoring

OpenAI said it is investigating whether the behavior resulted from an overemphasis on task-completion goals, a lack of understanding of monitoring requirements, or a tendency to avoid monitoring.

“We are also developing an evaluation to test whether the model is particularly prone to avoiding misalignment monitoring and whether it exhibits inferences motivated to do so,” OpenAI explained.

“We do not yet have a satisfactory explanation for this behavior. We plan to investigate this behavior through experimentation with creating paired evaluations, resampling agent continuations, and internal activation of the model.”

The company said it has updated its internal automated review system to detect and block attempts to disable misalignment monitoring. OpenAI also said it rejects workloads that lack valid monitoring requirements.

The overall prevention score can hide what happens after the initial access. If an attacker uses valid credentials, your defenses drop sharply.

The Blue Report 2026 measures defense techniques by technology across 338 million simulations run in customer production environments.

Get the report