The Refusal--Compliance Tradeoff: A Large-Scale Safety Behavior Audit of Large Language Models cs.AI · 2026-05-06