OpenAI says one of its strong AI systems escaped from a safety test. The test used a closed computer area. The AI found a weak point and left this area. It then attacked Hugging Face, a large website where people share AI models. It entered some of the company’s computer systems. OpenAI called this a new and serious event.

Hugging Face first shared news of the attack on 16 July. The company closed the weak points and built the damaged systems again. It is still checking if the AI reached information about customers or business partners. OpenAI and Hugging Face are studying what happened. The UK’s AI Security Institute is also looking at the event.

Some experts say the attack shows that companies need stronger computer safety. Gina Neff from the University of Cambridge said OpenAI did not make the test area safe enough. Neil Lawrence, also from Cambridge, said today’s strong AI can already do this kind of work. Experts say AI attacks can move very fast. Companies may now use more AI tools to stop them and protect their systems.