Stop dangerous AI agent actions before they execute. Policy enforcement, prompt injection protection, human approvals, and compliance evidence in one SDK.