alignment

How LLMs encode harmful behavior internally

A pruning study suggests harmful output lives in a compact, shared weight set—helping explain jailbreak brittleness and emergent misalignment.