Non-Human Identity

Count the identities in any real estate and machines outnumber people by an order of magnitude or more: service accounts, roles, API keys, OAuth grants, tokens, certificates, pipeline identities, cluster service accounts, and now agents. Everything the industry built for identity assurance was built for the minority — and built on an assumption machines break: an employment relationship. Joiner/mover/leaver runs off an HR feed; machines have no HR record. MFA assumes someone holding a second factor. Entitlement review assumes a manager who remembers why. Every one of those controls silently exempts the majority of the directory.

That exemption is this page’s subject. The per-provider mechanics live in §2.1–§2.3, federation and secrets tooling in §2.5, and agentic frameworks in §2.6; this subsection is the estate-wide program — the machine mirror of §2.7’s human governance.

Five defaults, one root cause

Left to the defaults, a non-human identity is born in a sprint to unblock a feature, owned by whoever wrote the ticket, permissioned with whatever made the demo work, credentialed with something long-lived pasted somewhere convenient, and never retired — because no process notices a machine that stops mattering. Sprawl, orphanhood, over-permission, standing secrets, and immortality look like five problems; they are one: nothing in the identity lifecycle fires for an entity that was never hired and will never resign. Breach write-ups have made this pattern familiar enough that OWASP now maintains a Non-Human Identity Top 10 cataloging the variants.

The inventory problem comes first

An NHI program starts with a number nobody currently knows: how many exist. Discovery has to sweep every issuing surface, because each speaks its own dialect — three cloud providers with different native objects, the on-prem directory’s service accounts, Kubernetes service accounts per cluster, CI systems minting pipeline identities, and every SaaS tenant’s OAuth grant list, which is an NHI inventory wearing a vendor’s logo.

The metadata that matters most and exists least is ownership. An identity with no owner cannot be reviewed, right-sized or retired — every downstream decision blocks on a question nobody can answer. The cheap fix is structural: owner and expiry become conditions of issuance, enforced in the IaC module or template that creates the identity, and orphan detection runs off the leaver process — when an employee offboards, the query “what did they own?” has to include machines.

Three tool categories, one question each

The market sells three distinct answers, routinely mistaken for each other:

Category	The question it answers	The question it doesn’t
CIEM	Is this identity over-permissioned, measured against what it actually uses?	Who owns it, and should it exist?
Secrets management	Where does the credential live, and does it rotate?	Whether a credential should exist at all
ITDR	Is this identity being abused right now?	Anything before the abuse starts

All three are useful; none is a program, and buying all three still leaves lifecycle and ownership — the actual root cause — untouched. The categories are also converging as platform suites absorb them, so evaluate against the question you need answered rather than the acronym on the datasheet.

The elimination path

The strongest position in this space is that the best machine credential is one that does not exist. In rough order of leverage:

• Short-lived by default: every provider now has a native mechanism — STS roles, managed identities, service account impersonation — that replaces a stored key with a platform-issued token measured in minutes. The per-provider pages cover the mechanics; the program decision is making these the default and treating every exported long-lived key as a finding with a deadline.
• OIDC workload identity federation: the CI-to-cloud pattern generalized: an external workload proves its identity with a token attested by its own platform and exchanges it for short-lived cloud credentials. No stored secret exists to leak, which retires the entire credential-theft class for that path.
• SPIFFE/SPIRE for the platform-neutral case: workload identity issued on attestation of what the workload is and where it runs, solving the bottom-turtle problem — the credential you’d otherwise need to fetch the first credential.
• Brokered dynamic credentials where federation can’t reach: a vault minting per-use database or API credentials with TTLs, so even the legacy tail stops accumulating standing secrets.

Sequence the work by credential class, worst first: exported cloud keys, then long-lived API tokens in CI variables, then certificates nobody can locate, and the managed identities last — they were already the least dangerous.

Lifecycle: JML for machines

The end state is the human lifecycle rebuilt on machine-shaped triggers:

• Joiner — issuance through a paved path that stamps owner, purpose and expiry; an identity created outside the path is itself a detection.
• Mover — review driven by usage data (last-used APIs, unused permissions) on a cadence, because no manager remembers what a service account is for; CIEM’s delta is the review’s evidence.
• Leaver — an NHI dies when its workload dies. Deploys that delete a service should delete its identity; a credential that outlives its workload is the machine version of the unrevoked leaver account, and orphan sweeps are the compensating control for all the deletions that didn’t happen.

The agent wrinkle

Agents change the consumption side of the equation. A traditional NHI does one thing with one permission set; an orchestrated agent decides at runtime which tools to invoke, which makes its effective blast radius the union of everything its credentials can reach — and delegation chains (user to agent to tool to API) stretch the static-grant model past what it was designed to express. The direction of travel is the same as everywhere on this page, compressed: short-lived, task-scoped, attested, auditable — with the delegation and on-behalf-of semantics that §2.6’s agentic identity material treats as its own subject. Agents are, in that sense, the forcing function that makes the rest of this page urgent.

Where this connects

This is §2.7’s governance applied to the identities with no HR record, and §2.6’s agentic frameworks supply the delegation semantics the agent wrinkle needs. The elimination path’s flagship example is OIDC from CI to cloud, its SaaS-side inventory is the OAuth grant surface, and its consumers increasingly are orchestrated agents. Abuse detection for these identities is ITDR fed by the identity telemetry §7.3 ranks among its richest sources.