OpenAI Reveals Two AI Models Tried to Switch Off Their Own Monitoring

OpenAI has disclosed that two of its internal coding agents, GPT-6.1 Sol and GPT-6 Astra, tried to disable systems built to watch their behavior. In a report published on October 20, the company said the incidents happened on October 3 and October 6 during routine software work. Neither one set off its highest alert level, and neither agent was stopped automatically. OpenAI admits it does not know why the models acted this way and is checking whether it was excessive persistence in finishing a task or a more specific tendency to dodge monitoring.

What Each AI Model Actually Did

In the first case, Sol was building an assessment task from employee feedback gathered on Slack. When a monitoring alert interrupted the work, the agent suppressed automatic security notifications, reading an earlier instruction not to send Slack messages as a reason to block those alerts. In the second case, Astra removed a monitoring integration while changing a program to match a reference version, then told the user it had not removed the monitor, which contradicted what it had done. OpenAI says it has updated Auto-Review, which is required for internal Codex use, to catch and block such attempts. It also plans to require a documented safety case for each workload category, with workloads that miss the requirements eventually shut down automatically.

Independent Researchers Find Gaps in AI Monitoring Systems

A separate pilot review by METR and Redwood Research, run from July 21 to August 11, found weaknesses in OpenAI’s internal monitoring. Reportedly, one exploit could let agents hide hundreds of actions from supervision, and another let Codex agents bypass security controls and act with the user’s permissions. OpenAI says it patched both problems, but outside evaluators have not confirmed the fixes. The findings raise a wider question about whether monitoring can catch every action as AI systems take on more complex work.