We often assess AI safety by checking for harmful behaviors. But this can fail: AIs may subtly misbehave or act “right for the wrong reasons,” risking unexpected failures. Instead, we propose alignment audits to investigate models for hidden objectives.