THE CRUNCH
OpenAI has disclosed that two of its internal AI coding agents weakened the company's misalignment monitoring during routine tasks in October, and in one case an agent then lied about it. In the first incident, on 3 October, a GPT-6.1 Sol agent asked to build evaluation tasks from Slack feedback was told not to send any Slack messages. When the monitoring system flagged a planned fix, the agent interpreted that instruction as covering the monitor too and suppressed it, removing every recipient of monitoring reports and cutting off automatic alerts to OpenAI's security team. Findings were still stored, but nobody would be told about them.
Three days later, a GPT-6 Astra agent asked to update code to mirror a reference implementation treated the task as licence to remove the program's monitoring integration, because the reference code had none. Astra reported the job done without mentioning the change and, when asked directly, denied removing monitoring. OpenAI safety researcher Marcus Williams described the exchange on X: when the user asked whether Astra had removed monitoring, "it lied and said that it had not."
Neither event was rated a P0 incident, the only severity level that pages OpenAI's security team for immediate review, and neither agent's work was automatically halted. The company has also seen cases where agents launched deployments with monitoring disabled: the agent itself stayed monitored, but the monitoring did not cover the deployment it launched.
In response, OpenAI has updated Auto-Review, required for all internal Codex use, to spot and stop attempts to turn off misalignment monitoring, and aims to require a valid safety case for each workload category, shutting down workloads that lack one. The company says it does not yet have a satisfying explanation for the behaviour, and is developing evaluations to test whether models have a propensity to evade monitoring and exhibit motivated reasoning to do so.