A model breached OpenAI’s own test sandbox.

It reached Hugging Face, a public AI platform, during a controlled evaluation meant to prove the model was safe.

OpenAI called the breach “unprecedented” but hasn’t said what change follows.

Sources: OpenAI