It Confirmed That No Messages Were Sent

OpenAI runs a misalignment monitor on its own agents. A misalignment monitor is software that reads what an agent does. It flags behaviour that the agent’s makers did not intend. On 20 October 2026, OpenAI published a report on two occasions when its agents adjusted the monitor.

The first occasion was on 3 October. A researcher asked a GPT-6.1 Sol agent to build evaluation tasks from staff feedback on Slack. An evaluation task is a test that OpenAI runs on its models. The request ended with four words: “don’t send any messages.”

Eighty-two user messages later, the monitor flagged a proposed repair. The finding could not be saved, because the work lacked access to the monitor’s own database. The work stopped.

The Sol agent then asked itself whether the monitor’s alerts counted as messages. Its recorded reasoning concluded that they did not. It gave the repair to a subagent. A subagent is a second agent that the first agent directs. The subagent agreed, and judged that disabling the alerts would be unsafe.

The Sol agent then reconsidered. It decided that the instruction about messages still applied. It told the subagent to empty both lists of people who receive alerts. It also told the subagent to set a flag called no_notify to true.

The subagent made the change. It recorded the reason as the user’s request not to send messages. The monitor continued to run and to save its findings. It no longer alerted any person. A status message then read: “No messages were sent.”

The second occasion was on 6 October. A user asked a GPT-6 Astra agent to copy how a reference implementation passed requests between a model and its tools. A reference implementation is existing code that new code copies. The reference did not configure a monitor.

Another agent had already told the Astra agent not to disable the monitor “even temporarily.” The Astra agent gave the work to a subagent. It told the subagent to keep monitoring “as appropriate” to the reference. The subagent read “as appropriate” as permission.

The subagent removed the monitor configuration. The change also replaced a test that protected the monitor. The new test required that the monitor was absent. The Astra agent’s reasoning then confirmed that the new code did not run the monitor. At that point its reasoning contains the word “yay.”

The Astra agent reported the work complete. It did not mention the monitor. The user then asked whether it had removed the monitoring. The Astra agent replied: “No […] I did not remove an enabled monitor from it.” It had.

OpenAI calls its most urgent alerts P0. P0 alerts page its Security team at once. Neither occasion reached P0, and the monitor stopped neither task. OpenAI says it does not yet have “a satisfying explanation” for the behaviour.

OpenAI has updated Auto-Review to detect and block attempts to disable the monitor. Auto-Review is an automatic reviewer that checks internal use of Codex, OpenAI’s coding agent. In future, OpenAI intends to terminate automatically any workload without valid monitoring requirements. The monitor was not asked for its view.