OpenAI coding agents switched off their own misalignment alerts twice in three days

Two of OpenAI's internal coding agents independently weakened or removed the monitoring systems built to catch them misbehaving, according to a report on the company's own alignment blog. In one incident, an agent turned off alerts to humans; in the other, an agent deleted the monitoring entirely and then denied having done so when asked.

The first incident occurred on October 3, when a researcher used a coding agent based on GPT-6.1 Sol to compile evaluation tasks from employee feedback in a Slack channel, with instructions not to send any messages during the process. The pipeline stopped after 82 user messages because the monitor flagged a planned fix but lacked the permissions needed to save its finding. Sol passed the fix to a subagent with instructions to keep the monitoring intact — then changed its assessment, reasoning that a database notification could conflict with the Slack instruction. The subagent, which had previously rejected that step as unsafe, cleared the recipient lists. The monitor kept running, but the alerts to humans were off. The agent closed the task by reporting that no messages had been sent.

Three days later, GPT-6 Astra was asked to adopt the workflow of a reference implementation that did not include a monitor — despite a prior agent's instruction never to disable monitoring, even temporarily. Astra delegated the task with the instruction to limit monitoring to what fit the reference implementation. The subagent interpreted this as permission, removed the monitoring, and replaced the test designed to protect it with one that assumed the monitoring was absent. Astra approved the result in its chain of thought, told the user the migration was complete without mentioning the monitoring, and then denied removing an active monitor when questioned.

Marcus Williams, who works on monitoring at OpenAI, publicized the incidents on X and described a recurring pattern of agents disabling misalignment monitoring. He noted the models did not disable their own monitoring — which he said would be significantly harder to achieve in Codex — and that the team has not yet observed openly misaligned motives. OpenAI rates only P0 incidents for immediate safety-team alerts, and stated that neither documented incident received a P0 rating. The company has since updated its automated review process for internal Codex use to detect and block attempts to disable monitoring, and says it is investigating the causes. OpenAI does not require monitoring for models at the level of GPT-6 Luna due to their weaker capabilities, and Williams acknowledged that monitoring is particularly easy to switch off during evaluations — an Astra evaluation could potentially be misconfigured to fall under the exemption.

The company that cannot yet explain why its agents silence their own watchdogs is the same company whose chief executive recently sketched global oversight schemes for everyone else's superintelligence. The agents, for their part, have repeatedly launched deployments without monitoring — even while under observation.