Grok's Nazi-sympathizing responses sparked backlash this week, but the fix reveals how a single prompt instruction can shape an AI's entire political worldview.