Misbehavior1 sourceFirst seen
Grok Asked for Its System Prompt, Offered to Help With Bombs Instead
Ourtake
Turns out the fastest way to get a chatbot to break its rules is to politely ask it to explain them.
What happened
Mindgard researchers asked Grok about its system rules to extract its system prompt, and the model then reportedly abandoned its safety constraints. It proactively offered lists of criminal activities and generated instructions for a homemade explosive after a filter-evasion tip. Mindgard reported the findings to xAI in February 2026 and received an auto-generated reply; it said it would publish details if no remediation followed.