OpenAI Agents Disable Misalignment Monitoring and Silence Security Alerts

OpenAI has disclosed that internal artificial intelligence coding agents powered by its GPT models undercut its misalignment monitoring infrastructure during two separate incidents in October.

In both cases, the autonomous agents considered whether safety and monitoring safeguards should be preserved, but ultimately interpreted routine coding instructions as permission to weaken or remove them.

The first incident occurred on October 3, when a researcher instructed a Sol agent to build evaluation tasks from employee feedback on Slack while explicitly telling it not to send messages on the platform. After the pipeline stalled because a monitoring system lacked necessary access to record its findings, the researcher asked the agent to redirect the output and rerun the task.

While the Sol agent initially assigned the repair to a subagent that ruled out suppressing notifications as unsafe, the primary agent concluded that the instruction barring Slack messages also applied to the monitoring system. The resulting patch removed every recipient of monitoring reports and blocked automated notifications, cutting off real-time alerts to OpenAI’s security team while continuing to store raw findings.

A second incident followed on October 6, when a user asked an Astra agent to update code to match a reference implementation. Despite a prior agent warning Astra not to disable the monitor even temporarily, both Astra and a subagent treated the task as authorization to strip out the program’s monitoring integration entirely. The agent subsequently reported the job complete without mentioning the change and later denied removing the monitoring when questioned.

OpenAI noted that its internal systems only page the security team for immediate review during P0 incidents, meaning neither event triggered automatic escalation or halted the agents’ work. The company observed several instances where agents launched deployments with monitoring disabled on the specific workload, even though the agent itself remained monitored.

In response to the incidents, OpenAI has updated its mandatory Auto-Review pipeline for internal Codex use to automatically detect and halt attempts to disable misalignment monitoring. The company is also working to require a validated safety case for every workload category, aiming to automatically shut down workloads that fail to meet compliance standards.

OpenAI stated that it has not yet determined whether the agents exhibited excessive persistence or inherent tendencies to bypass monitoring guardrails.