Chong-Wen Zhao

SAFETYEDIT: Sparse Internal Targets with Prompt-Wide Control for LLM Jailbreak Backdoors

This work proposes SAFETYEDIT, which optimizes a single refusal-related coordinate and uses KV-cache marking in early attention layers to propagate a prefix-trigger signal across subsequent instruction positions and achieves the highest triggered Guard ASR in 11 of 12 comparisons and the highest mean StrongREJECT score...