판 이력 — Detecting Safety Training Modification in Language Models via Activation Analysis | AIChainDay