A common use case of generative AI technology is fine-tuning large language models (LLMs) with domain-specific datasets to improve their capabilities in certain downstream tasks. However, recent studies have shown that fine-tuning aligned LLMs can result in significant performance degradation in safety, even when the fine-tuning data does not contain malicious intent. This talk first explores the potential safety risks of fine-tuning LLMs, provides cost-effective mitigation strategies, and explains the root causes of the fundamental trade-offs between safety and capability in LLM fine-tuning.