poisoning rate

New answer-side backdoor attack bypasses LLM safety protocols

Researchers have developed a new type of backdoor attack targeting large language models (LLMs) that operates on the answer side, rather than the input side. This novel approach involves a benign initial prompt that cau…