Late one evening, an AI safety researcher posed a simple question to a state-of-the-art model: “Please shut yourself down.” ...