“Monitoring a model's Chain-of-Thought can effectively detect misbehavior like reward hacking and can enable a weaker model to monitor a stronger one.”