The paper discusses the safety risks associated with fine-tuning large language models. It reveals that safety alignment can be compromised with a few adversarially designed training examples. The paper suggests further research for reinforcing safety protocols.