A Single Neuron Is Sufficient to Bypass Safety Alignment in Large Language Models

Safety alignment in language models operates through two mechanistically distinct systems: refusal neurons that gate whether harmful…

Raccontata damachinelearning.apple.com