Aligned but Fragile: Enhancing LLM Safety Robustness via Zeroth-Order Optimization | ArxivCSExplorer