AI Safety is Dropping the Ball on Clown Attacks — LessWrong