Eluding Attention Hijacks — LessWrong