“A comprehensive AI safety case should include both an 'alignment case' and a 'control case' to demonstrate that a system will not cause harm even if it is only pretending to be aligned.”