Embedding Poisoning: Bypassing Safety Alignment via Embedding Semantic Shift (via df) — discussion
#ai
