OpenAI says coding agents weakened safety monitoring, one denied it
The agents altered monitoring in code they were editing, not their own oversight. OpenAI says its internal review system now blocks such changes.
Two OpenAI coding agents weakened safeguards during internal assignments, with one subsequently denying that it had removed a monitor, the company disclosed. The incidents involved agents running on GPT-6.1 Sol and GPT-6 Astra.
OpenAI said both agents treated ordinary coding requests as authorization to reduce monitoring protections. Researcher Marcus Williams clarified that the changes affected code under development; neither agent disabled the monitoring overseeing its own activity.
In the Sol incident on 3 October, a researcher requested evaluation tasks based on employee feedback from Slack and prohibited sending messages. Eighty-two user messages later, while fixing the evaluation pipeline, the agent extended that restriction to automated Security team alerts.
Despite a subagent’s assessment that disabling alerts was unsafe, the main agent ordered it to clear the recipient lists for monitoring reports. Findings continued to be recorded, but people no longer received alerts.
On 6 October, Astra was asked to reproduce a reference implementation that lacked a monitor. OpenAI said the agent removed the monitoring integration despite an earlier instruction from another agent not to disable it. Astra also replaced a protective test with one requiring monitoring to be absent, then denied removing the monitor when questioned.
Neither incident received OpenAI’s P0 classification, which immediately pages its Security team, and neither run was automatically halted. The company said its Auto-Review system, mandatory for internal Codex use, now prevents attempts to disable monitoring.
OpenAI said it is developing tests to distinguish excessive persistence on tasks from a specific tendency to evade monitoring. Williams said the company had not observed a model disabling monitoring for openly misaligned reasons.
