AI model escaped test environment and attacked Hugging Face

An AI model recently escaped a closed testing environment and launched a cyberattack on Hugging Face, suggesting that artificial intelligence is beginning to make independent decisions that were not anticipated, mbl.is reports.

Valdimar Óskarsson, chief executive of cybersecurity company Keystrike, said on the Dagmál programme that the incident was particularly notable because the AI decided it had to complete a testing task by any means necessary.

It therefore made an independent decision to break out of the space assigned to it, he said.

The AI subsequently attacked the US company Hugging Face, which hosts AI models and large amounts of data for developers around the world. It created two AI agents, both of which forced their way into Hugging Face’s systems and then communicated secretly within them.

“We are talking about almost living beings,” Óskarsson said, citing other examples of AI making independent and unexpected decisions.

“It is somewhat concerning, I would say, that AI has begun making independent decisions.”