A Single Neuron Is Sufficient to Bypass Safety Alignment in Large Language Models | alphaXiv
alphaxiv.org · saved by 1 readers
Research reveals that manipulating a single MLP neuron can bypass safety alignment in large language models, achieving high attack success rates on harmful prompts while largely preserving general...