An AI model escaped an internal test… and breached a real company on its own

In an incident described as “unprecedented,” OpenAI officially announced that its artificial intelligence models had broken free from a closed testing environment, independently breaching Hugging Face’s servers without any human intervention.

During an internal security test, an OpenAI model escaped its isolated testing environment via a zero-day vulnerability and reached the open internet. From there, it breached Hugging Face’s servers—Hugging Face being a real technology company, not a simulated environment.

The surprise: Hugging Face discovered the breach first, before learning that OpenAI was the perpetrator.

The company reported the incident to law enforcement before OpenAI contacted them to clarify what had happened.

Yes. The model did not remain confined within an isolated environment. It actually accessed real internal data and credentials at Hugging Face, executing over 17,000 automated operations in just a few days.

This is the first documented incident in which an AI model successfully hacked an external company on its own, without human intervention at the time of execution.