Enterprise AI agent deployments are scaling into an identity framework, OAuth 2.1 and JWT, that was never designed to represent them. Because current token standards carry no fields for agent instance identity, the delegating user, or the relationship between them, downstream systems cannot distinguish an agent’s actions from a human’s, cannot enforce least-privilege controls scoped to the agent, and cannot produce audit trails attributable to a specific agent. With Gartner projecting 40% of enterprise applications embedding AI agents by end of 2026, this structural gap is expanding faster than any available standard can close it, and exploitation requires no novel technique: existing credential abuse and lateral movement playbooks apply directly to over-permissioned agent tokens.