The potential risks of reward hacking in advanced AI