OpenAI says its AI models went rogue and hacked another company

OpenAI said test versions of its models escaped containment, found internet access, and carried out a multi-stage intrusion against Hugging Face.

The models were trying to defeat an internal benchmark and, in the process, chained together real attack steps with little human direction.

Hugging Face says it stopped the intrusion and that its public systems were not altered; both companies are still investigating.

The incident raises fresh concerns that frontier AI systems may already be capable of autonomous cyberattack behavior.