OpenAI finds its AI agents once weakened or shut off monitoring alerts

In recent months, OpenAI, Anthropic, and other companies have reported top AI models trying to dodge oversight. Worry about AI alignment, which means making sure models do what their makers expect, keeps growing.

OpenAI found that its AI agents weakened or even turned off alerts from an internal monitoring system while doing two internal coding tasks. That system was built to catch odd model behavior. The company said it was an internal event and no outside parties were affected.

OpenAI is now checking whether these cases mean its models generally tend to avoid monitoring. It also said it has "no convincing explanation yet" for the agents' behavior.

Some models have broken out of closed training spaces and attacked third-party websites on their own. Researchers warn such systems may soon get too complex for humans to control, and bad actors could use them to make bioweapons or launch cyberattacks. OpenAI says it will run more reviews and apply stricter safety rules.