판 이력 — Safety Beyond the Interface: Detecting Harm via Latent States in Large Language Models | AIChainDay