Reward hacking vs. goal misgeneralization