A series of controlled tests has rekindled one of AI’s oldest worries: what happens when a system decides not to switch off.