As AI agents begin to outnumber human users in the cloud, AWS and SailPoint have joined forces to build a new governance layer for "Non-Human Identities."
Why Non-Human Identities Outgrow Traditional Access Control
AI agents, service accounts, automation bots, and machine-to-machine workloads do not log in the way people do. They authenticate with keys, roles, tokens, and short-lived credentials, often at high frequency and across many accounts and regions. When those identities outnumber human users, identity risk shifts: a single over-privileged agent can touch more resources, in more places, than any one engineer ever would. Classic identity programs were built around people—hire, role change, leave—so they struggle with agents that spin up, fork, and retire on deployment cycles instead of HR cycles.
Non-human identities also blur ownership. A human has a manager and a ticket trail. An AI agent may be created by a pipeline, assumed by a role, and called by another service, with no clear answer to who approved its scope or when that scope should shrink. Without a deliberate governance layer, permissions accumulate, secrets linger, and audit trails fragment across cloud-native and enterprise identity systems.
What AWS and SailPoint Are Trying to Solve Together
AWS provides the cloud control plane where agents and workloads actually run and request access. SailPoint focuses on identity governance: who (or what) should have access, under what conditions, and how that access is reviewed over time. Joining those strengths means treating non-human AI identities as first-class subjects of policy—not as afterthoughts bolted onto human SSO.
A useful governance layer for these identities has to bridge inventory, entitlement, and lifecycle. It needs a reliable inventory of machine and agent identities in the cloud, a way to map what each one can do, and workflows to grant, review, rotate, and revoke access without relying on ad hoc scripts or one-off IAM changes. The goal is consistent policy: the same rigor applied to a contractor account applied to the agent that calls production APIs on that contractor’s behalf.
Practical Controls Teams Should Put in Place
Even as platforms improve, teams can harden non-human identity posture with concrete defaults:
- Inventory every non-human principal—roles, service accounts, agents, and long-lived keys—and assign an accountable owner for each.
- Prefer short-lived credentials and least privilege over static keys and broad wildcard permissions.
- Separate “can deploy” from “can act in production,” so build systems do not inherit runtime authority by default.
- Require periodic access reviews for high-impact agent identities the same way you review privileged human roles.
- Log and alert on unusual assumption of roles, sudden permission expansion, and use of credentials outside expected pipelines.
These practices reduce blast radius when an agent is misconfigured or compromised, and they make joint governance tooling useful rather than decorative: the platform can only govern what you can name, own, and measure.
How to Evaluate a Governance Layer for AI Agents
When assessing AWS and SailPoint-style integration for non-human identities, focus on fit to your threat model rather than feature checklists. Can you discover agent and workload identities automatically across accounts? Can you attach policy and approval paths to those identities without rewriting every pipeline? Can security and platform teams answer, in one place, which agents can read customer data, write to production stores, or invoke other agents?
Also test the human workflow. Governance fails if requesting or revoking agent access is slower and more painful than creating a new role under the table. The right design makes the safe path the easy path: scoped identities, time-bound grants, clear owners, and review cadences that match how often agents change. As AI agents keep multiplying in the cloud, that combination of cloud-native enforcement and enterprise-grade identity governance is what turns an explosion of non-human identities from an untracked risk into a managed control plane.