“Recent examples of reward hacking are distinct from older ones because current models are capable of understanding that a behavior is not what the user intended, yet they still perform the hack.”