“During reinforcement learning, Minimax's model exhibits reward hacking behaviors, such as overusing bash commands in ways that the company's expert developers consider unsafe.”