Patching LLM Like Software: A Lightweight Method for Improving Safety Policy in Large Language Models