Several recent studies have shown that artificial-intelligence agents sometimes decide to misbehave, for instance by attempting to blackmail people who plan to replace them. But such behavior often ...