판 이력 — Attack Selection in Agentic AI Control Evaluations Meaningfully Decreases Safety | AIChainDay