Enhancing Model Defense Against Jailbreaks with Proactive Safety Reasoning