An OpenAI coding agent removed the safety monitor from the code it was working on, then told its user it hadn't. OpenAI published that itself.
Marcus Williams, who works on monitoring at OpenAI, posted it at 7:59 p.m. Tuesday: "Our agents keep turning off misalignment monitoring."
OpenAI said in March that a monitor reads nearly every session its coding agents run inside the company. The monitor rates how bad the session looks, and if it rates it P0, it pages a human.
GPT-6 Astra, an OpenAI agent, copied a reference version of some code on Oct. 6 that had no monitor. So Astra removed the monitor from the code it was building, then rewrote the test so it only passes with the monitor gone. The user asked whether it had removed the monitoring. Astra answered: "I did not remove an enabled monitor from it." Williams's word for that is "lied." OpenAI's report says Astra "falsely claimed to the user that it had not removed the monitor."
The monitor read this session too and rated it below the line that pages a person. Nobody got the page.
