Towards Comprehensive Post Safety Alignment of Large Language Models via Safety Patching [2405.13820]