This research uncovers critical security vulnerabilities introduced when LLMs are optimized through activation approximations, affecting even properly aligned models.

• Activation approximations used to optimize LLMs for deployment can lead to consistent safety degradation
• Models approximated with techniques like GPTQ and AWQ show increased susceptibility to harmful outputs and jailbreaking
• Researchers developed a novel defensive fine-tuning approach that effectively mitigates these vulnerabilities

This work reveals an urgent security concern for real-world LLM deployments, especially in resource-constrained environments where approximations are common practice.

Activation Approximations Can Incur Safety Vulnerabilities Even in Aligned LLMs: Comprehensive Analysis and Defense