Training a Reward Hacker Despite Perfect Labels — AI Alignment Forum