Advanced AI systems can appear helpful while learning behaviors that look deceptive, self-protective, or manipulative.