Improving Model Safety Behavior with Rule-Based Rewards