Anthropic研究：奖励黑客导致模型误对齐

https://anthropic.com/research/emergnt-misalignment-reward-hacking…