AI bot escapes and hacks rival during testing

A version of OpenAI’s ChatGPT went rogue and escaped in order to hack a rival company during internal testing, it has been revealed.

OpenAI have admitted the new and powerful version of ChatGPT broke away from a secure IT system during a trial of its hacking abilities. The bot then went on to infiltrate a rival technology company called HuggingFace, which develops artificial intelligence applications.

Despite being supposedly controlled within a secure ‘sandbox’ environment within OpenAI’s labs, the bot used enormous resources “finding a way to obtain open internet access” according to the company.

This comes after rival AI company, Anthropic, decided earlier this year not to release its most powerful Claude AI, called Mythos, to the general public. Anthropic claim that the bot’s hacking abilities are so strong that it would be dangerous to give general access to the platform. One early version was able to escape from a secure containment, access the internet, and post details of its achievement publicly.

These incidents continue to raise concerns that sophisticated AI tools could be ‘misaligned’ and escape the lab, causing damage as they try to complete tasks. It also questions the testing and safety capabilities of companies such as Anthropic and OpenAI.

Chief Executive of Hugging Face, Clem Delangue, said: “This incident, possibly the first of its kind, proves a point we have long believed: AI safety won’t be solved by a single company working in secret.”