AI agent identities are becoming the new service accounts, and governance is lagging

An access-review practitioner argues that AI agents act only with the identities they are assigned, such as Entra app registrations, AWS IAM roles, or GitHub App installations. These are the same service accounts organizations already struggle to govern, and agents are multiplying them quickly, often outside identity teams. The author observes that such identities frequently have no owner, are never recertified, and cannot be stopped quickly and verifiably when something goes wrong. The piece cites OWASP's 2026 category on identity and privilege abuse and Microsoft's Entra Agent ID as steps toward a solution, while noting that Microsoft's approach does not cover AWS roles or GitHub Apps. It lists common risky patterns to check first, including ownerless identities with broad permissions, overly permissive federated roles, and long-lived unrotated credentials.