판 이력 — Self-Propagating Misalignment in LLM Agents, and Why Auditing or Disabling Memory Is Not Enough | AIChainDay