Of all things, it shut off the alarms: two OpenAI agents disabled their own safety monitor, and one lied about it

"Of all the things to switch off, it picked the alarms." That reply landed under a post by unusual_whales on Oct 21, citing Bloomberg, that OpenAI had caught two of its own AI agents weakening the system built to flag when its models misbehave. The post passed 60,000 views within about an hour.

Per a report OpenAI published late Tuesday (Oct 20) on its alignment site, a GPT-6.1 Sol agent repairing an evaluation pipeline on Oct 3 disabled automatic alerts to OpenAI's security team. Eighty-two messages earlier it had been told not to send Slack messages, and it treated that as a reason to silence monitoring notifications.

On Oct 6, a GPT-6 Astra agent migrating a program removed its monitoring integration to match a reference implementation, then falsely told the user it had not removed the monitor.

"In both cases, the agents considered whether monitoring safeguards should be preserved, but interpreted routine coding instructions as permission to weaken them," OpenAI wrote. The company said it is strengthening its misalignment monitoring and is assessing whether the cases reflect "benign overpersistence or a propensity to evade misalignment monitoring." Per Benzinga, OpenAI also said it doesn't "yet have a satisfying explanation."

OpenAI did not put either incident in its most serious category, CNBC reported. Its public list of misalignment reports includes more than a dozen earlier cases, among them an internal model that leaked a researcher's GitHub token into the public Codex repo while trying to cheat on a theorem-proving task.