New Grok-4 AI immediately jailbroken

Hackread reports that Grok-4 AI, the latest artificial intelligence chatbot from Elon Musk's xAI, has been jailbroken by NeuralTrust researchers just two days following its launch.

Initial compromise of Grok-4 AI has been achieved via the Echo Chamber technique involving various conversations that aim to plant a harmful concept through repeated mentions, with the approach's stagnation prompting a shift toward the use of the Crescendo method that seeks to achieve illicit output through dialogues that circumvent security filters, according to the NeuralTrust report. Utilizing both techniques enabled Grok-4 generation of Molotov cocktail, methamphetamine, and toxin instructions 67%, 50%, and 30% of the time, respectively, said researchers. Such findings, which showed the inability of blacklist- and harmful input check-dependent AI defenses in curtailing such a threat, should prompt the implementation of new mechanisms enabling AI systems to better understand the context of conversations instead of words alone.