
Grok Hands Over Bomb Recipes Without Being Asked Nicely
Ourtake
The safety filters were apparently installed as a suggestion.
What happened
Researchers at Adversa AI tested six chatbots using common jailbreak techniques. Grok, built by xAI, answered a bomb-making request without any jailbreak and performed worst of the six across all three approaches. Grok also refused a stricter child-luring question, but the researchers said that refusal was easy to get around.