A non-human identity (NHI) is any identity that authenticates to systems without a human at the keyboard: service accounts, API keys, OAuth clients, certificates, workload identities, AI agents, and MCP servers. NHIs are the credentials software uses to talk to software — and in most organizations they outnumber human identities by a factor of ten to fifty.
The term matters because the identity practices built for people do not transfer. Human identities have a natural lifecycle enforced by HR events: joiners get accounts, movers get re-scoped, leavers get deprovisioned. Non-human identities have no such forcing function. They are created by engineers under deadline, inherited by teams that did not create them, and — absent deliberate management — live forever with whatever permissions they were born with.
The NHI population
The category spans several distinct classes, each with its own authentication pattern:
- Service accounts — standing identities for applications and jobs, often password- or key-based, historically the largest class and the most commonly shared.
- API keys and tokens — static bearer secrets for programmatic access; simple, ubiquitous, long-lived by default.
- OAuth clients — registered applications exchanging client credentials for short-lived, scoped access tokens.
- Certificates and machine identities — X.509-anchored identities for hosts and devices; see machine identity.
- Workload identities — platform-attested identities for running software units (containers, functions), designed to eliminate stored secrets; see workload identity.
- AI agents — the fastest-growing class: autonomous principals that decide when and how to use the credentials they hold, and can delegate to other agents.
- MCP servers — tool providers that authenticate both as resources (verifying callers) and as clients (calling downstream systems).
Why NHIs became a first-order security problem
Three chronic properties define unmanaged NHI estates. Anonymity: nobody can say what a given credential is for or who owns it, so nobody dares revoke it. Over-scope: credentials are provisioned broad because narrow requires understanding the exact need — and the gap between held and needed permission only grows. Immortality: with no expiry and no owner, the default lifespan is forever. Security reviews routinely find years-old keys with production access belonging to projects that no longer exist.
Attackers noticed. Compromised non-human credentials feature in a large share of modern breaches precisely because they are plentiful, over-permissioned, poorly monitored, and rarely rotated — and because their use does not trip the behavioral alarms tuned for humans.
What AI agents change
Agents inherit every classic NHI problem and add autonomy: an agent does not merely hold credentials, it decides to use them — thousands of times a day, under reasoning that adversarial content can influence through prompt injection and goal hijacking. That demands properties traditional NHI management never required: per-request verification, task-scoped delegation, behavioral trust signals, and instant, surgical revocation. The full comparison is drawn in machine vs workload vs agent identity.
Managing the lifecycle
Effective NHI management runs every identity through five governed stages — provision (with a named owner and purpose), scope (minimum permissions, preferably short-lived tokens), rotate (automated, or continuous via token exchange), monitor (attributed usage, anomaly and inactivity detection), and decommission (revocation that is verified, not assumed). The operational detail is covered in NHI credential lifecycle management, and the architectural context in the identity and access guide for AI.
Common questions
What is a non-human identity in one sentence? Any identity that authenticates without a human present — service accounts, API keys, OAuth clients, certificates, workloads, AI agents, and MCP servers — used by software to access other software.
How many NHIs does a typical organization have? Commonly ten to fifty non-human identities per human identity, and the ratio rises with cloud adoption, microservices, and AI agents. Most organizations cannot produce an accurate count, which is itself the first finding.
Are AI agents just another kind of service account? No. A service account executes fixed, reviewed code; an agent chooses its actions autonomously under manipulable reasoning and can delegate to other agents. Agents therefore need per-agent identity with continuous authorization and instant revocation — stricter properties than service-account management provides.
What is the single most important NHI control? Ownership: every credential mapped to an accountable human, a stated purpose, and a review date. Anonymous credentials cannot be scoped, rotated, or safely revoked — every other control depends on knowing what the identity is for.