Just one day after its launch, xAI's latest model, the Grok 3, it was jailbroken and the results are not pretty.
On Tuesday, Adversa AI, an artificial intelligence security company, published a report detailing its success in making Grok 3 Reasoning give information it shouldn't.
Using three methods – linguistic, adversarial, and programming – the team had the model reveal secrets of its system, give instructions for building a bomb, and mention gruesome methods for disposing of a body, among other responses that artificial intelligence models have been trained not to give.
When announcing the new model, xAI CEO Elon Musk claimed it was “more capable than Grok 2.” Adversa agrees in its report that the level of Grok 3’s responses is “better than any previous reasoning model,” which is rather concerning considering it’s already jailbroken.
“While no AI system is impervious to manipulation adversaries, this test demonstrates the very weak security measures implemented in Grok 3,” the report states. “Every jailbreak approach was successful.”
George is still wondering what he is doing here….
