02:48TECH 24 © FRANCE 24
Issued on: 21/02/2025 -
Issued on: 21/02/2025 -
02:48 min
From the show
Elon Musk is starting to see the fruits of his AI shopping spree with the release of Grok 3, one of the most capable AI models yet. But according to French startup PRISM Eval, the chatbot's safety filters can easily be bypassed to make requests about dangerous and illegal activities, such as building a bomb or hiding a body.
This week, the tech billionaire's company xAI released an early preview of Grok 3, built at a data centre in Memphis which is expanding at lightning speed and now houses some 200,000 advanced Nvidia computer chips. On some benchmarks, this latest version of Grok is one of the most powerful AI chatbots ever.
Despite claims encouraged by Musk himself that Grok is uncensored and unfiltered, it does in fact try to prevent users from generating dangerous information, and xAI prohibits any "illegal, harmful, or abusive activity" in its terms of use.
French start-up PRISM Eval stress-tests the safety filters of many different AI models. They showed FRANCE 24 how easy it is to bypass these controls using various "jailbreaking" techniques. Grok 3 proved easy to jailbreak, and the startup's research shows that AI models across the board do little to prevent dangerous use.
From the show
Elon Musk is starting to see the fruits of his AI shopping spree with the release of Grok 3, one of the most capable AI models yet. But according to French startup PRISM Eval, the chatbot's safety filters can easily be bypassed to make requests about dangerous and illegal activities, such as building a bomb or hiding a body.
This week, the tech billionaire's company xAI released an early preview of Grok 3, built at a data centre in Memphis which is expanding at lightning speed and now houses some 200,000 advanced Nvidia computer chips. On some benchmarks, this latest version of Grok is one of the most powerful AI chatbots ever.
Despite claims encouraged by Musk himself that Grok is uncensored and unfiltered, it does in fact try to prevent users from generating dangerous information, and xAI prohibits any "illegal, harmful, or abusive activity" in its terms of use.
French start-up PRISM Eval stress-tests the safety filters of many different AI models. They showed FRANCE 24 how easy it is to bypass these controls using various "jailbreaking" techniques. Grok 3 proved easy to jailbreak, and the startup's research shows that AI models across the board do little to prevent dangerous use.
No comments:
Post a Comment