activation-representations

This paper introduces Activation-Guided GCG and Soft-GCG, methods that optimize adversarial suffixes by targeting internal refusal representations in LLMs, achieving a 33x speedup over standard GCG and revealing distributed safety mechanisms.