Measures that limit what an AI system is able to do, so that it cannot cause harm even if its goals are flawed.