판 이력 — Safe at One Loop, Risky at Another: Aligning Safety Across Recurrent Depths in Looped Language Models | AIChainDay