This research reveals how security mechanisms work inside aligned language models, identifying specific layers that serve as security gatekeepers.

• Discovered safety layers: A small set of contiguous layers in the middle of the model crucial for security
• Identified security vulnerability: These layers can degrade during fine-tuning
• Developed SPPFT: A technique to prevent security degradation while maintaining model performance
• Demonstrated a practical approach to strengthening LLM security at the parameter level

This work is significant for the security community as it provides a deeper understanding of how aligned LLMs resist harmful prompts and offers concrete methods to protect these mechanisms during model adaptation.

Safety Layers in Aligned Large Language Models: The Key to LLM Security