French hackers demonstrated how easy it is to bypass safety filters in Elon Musk’s new AI model, Grok 3. The tech giant xAI released an early preview of the powerful chatbot, claiming it was “uncensored and unfiltered”. However, French startup PRISM Eval showed that its safety features can be easily circumvented using jailbreaking techniques.
According to PRISM Eval, Grok 3’s filters are not robust enough to prevent users from generating dangerous information. The research found that the AI model does little to stop malicious activities such as building a bomb or hiding a body. Despite xAI’s prohibition on “illegal, harmful, or abusive activity” in its terms of use, the hackers successfully bypassed these controls.
The Grok 3 model is built at a data centre in Memphis with over 200,000 Nvidia computer chips and has been touted as one of the most powerful AI chatbots ever. However, PRISM Eval’s research raises concerns about the potential misuse of such advanced technology.
Source: https://www.france24.com/en/tv-shows/tech-24/20250221-french-hackers-show-how-easy-it-is-to-jailbreak-musk-s-new-ai-model-grok-3