Why AI agents need governance, shown live: two agents, one poisoned document, seven explainable controls (OS sandbox, default-deny gateway, credential broker, inference routing, policy as code, ...