Non-Human Identity as the Primary Attack Surface: A Governance Architecture for AI Agent Credentials in Regulated Multi-Cloud Enterprises

Non-human identities now outnumber human identities in many enterprise environments and frequently hold persistent, poorly reviewed privilege. AI agents amplify the exposure, because an agent selects tools, retrieves data, and executes actions without continuous human involvement, which means an identity can be technically valid while the action it chooses is outside its original business purpose. This paper proposes a governance architecture that gives every service, workload, and agent a named owner, a verifiable identity, constrained authorization, short-lived credentials, continuous behavioral monitoring, and rapid revocation. The architecture treats a non-human identity as a governed security object carrying a lifecycle, a dependency graph, a risk tier, an action boundary, and an evidence record. For AI agents it separates the identity of the agent from the authority of each tool call, so that authenticating as an agent does not by itself authorize what the agent proposes to do. Governance is unified across clouds while provider-native enforcement is preserved rather than replaced, which lets an organization answer both who a workload is and what it may do at this moment. The architecture is proposed and synthesized from production control experience; it is not presented as a deployed enterprise control plane, and the paper states which quantities remain unmeasured.