The Hard Authorization Boundary: Preventing Unauthorized Actions in Agentic AI