Definition
A non-human identity is an identity assigned to software, or to a device acting as a workload, and not to a person. It is used to authenticate to or be authorized by another system. It is not "anything that is not human". A laptop, a model and an API key are not identities: the first is a device, the second is software, the third is a credential.
This is an NHI Governance model, version 1.0, published 7 October 2026. It is our own analytical model. It is not a standard, and no standards body has endorsed it. Reuse is permitted under CC BY 4.0 with credit.
Three things that get confused
Much confusion in this field comes from using one word for three different things. An actor is the software that does something. An identity is how a system knows that actor. A credential is what the actor presents to prove the identity. Governance attaches to the identity. Rotation attaches to the credential. Accountability attaches to a person.
Fifteen terms, defined against each other
Scroll sideways to see all columns.
| Term | Definition | How it differs from its neighbours |
|---|---|---|
| Actor | The thing that does something: a person, or a piece of software such as a workload, service, script or AI agent. | An actor is not an identity. One actor can hold several identities, and several actors can share one, which is a governance problem. |
| Workload | A running unit of software: a container, function, virtual machine or batch job. | A workload is an actor. It needs an identity to call anything else. |
| Application or service | Software that provides a function to users or to other software. | It usually runs as one or more workloads and is what an owner recognises by name. |
| Machine | A device or host. In older usage, any non-human actor. | In precise use we mean devices and hosts. In everyday phrases such as machine identity or machine action the word covers any non-human actor; where it matters we say workload, service or device. |
| AI agent | Software that uses a model to choose which actions to take toward a goal, within authority someone delegated to it. | An agent is an actor. What makes it different is that its next action is not fixed in code. |
| Identity | The record by which a system knows an actor: an account, principal, role or attested name, with attributes. | An identity is governed: it has an owner, a purpose and a lifecycle. It is not the secret that proves it. |
| Non-human identity (NHI) | An identity assigned to software, or to a device acting as a workload, and not to a person. It is used to authenticate to or be authorized by another system. | The test is who the identity is assigned to, not what uses it at a given moment. A person using a service account is misuse of an NHI, not a human identity. |
| Credential | The material an actor presents to prove an identity: a password, key, token or certificate. | Credentials are rotated, expired and revoked. One identity can have many credentials over its life. |
| Secret | A credential that works for anyone who holds it and must therefore be kept confidential. | All secrets are credentials. Not all credentials are secrets: a certificate is public, its private key is the secret. |
| Authentication | Establishing that an actor is the identity it claims to be. | It answers who is calling. It says nothing about what the caller may do. |
| Authorization | Deciding whether an authenticated identity may perform a specific action on a specific resource. | It answers what is allowed. Most excessive privilege is an authorization problem, not an authentication one. |
| Permission | A specific grant: a role, scope, policy statement or entitlement attached to an identity. | Permissions are what an access review examines. |
| Action | One operation performed by an actor using an identity: an API call, a query, a write. | Actions are what logs record and what an incident is made of. |
| Evidence | A record that can show a third party what happened and under whose authority: logs, approvals, review decisions, configuration history. | Evidence is produced by the programme, not asserted by it. |
| Accountability | The obligation of a named person to answer for what an identity was allowed to do and did. | Accountability cannot be assigned to software. Every non-human identity needs a human who holds it. |
Eleven identity types
These are the kinds of non-human identity an enterprise typically holds. The credential column shows what proves each one. The same credential type can serve several identity types, which is why counting keys is not the same as counting identities.
Delegation is not an identity type. When an agent or service acts on behalf of a person or another agent, it uses one of the identities below with authority that was passed to it. See delegation chains.
Scroll sideways to see all columns.
| Identity type | What it is | Typical credential | Where it lives |
|---|---|---|---|
| Service account | Account in a directory, database or platform, created for software | Password, key or keytab; on Kubernetes, a projected service account token | Active Directory, databases, SaaS, Kubernetes |
| Service principal or application registration | Identity of an application in a cloud identity provider | Client secret, certificate, federated credential | Microsoft Entra ID, other identity providers |
| Cloud role or managed identity | Role or identity a workload assumes on a cloud platform | Short-lived platform-issued token | AWS IAM, Azure, Google Cloud |
| Workload identity | Identity bound to a workload by attestation, not by a stored secret | Short-lived signed document or certificate | SPIFFE, Kubernetes, cloud federation |
| OAuth client or application | Third-party or internal application granted delegated access | Client credential, access token, refresh token | SaaS platforms, identity providers |
| API consumer | Caller identified to an API by a key | API key | Any API |
| CI/CD identity | Identity a pipeline run uses to deploy or to reach other systems | Stored secret or federated token | GitHub Actions, GitLab, Jenkins |
| Certificate subject | Host, service or device named in a certificate | X.509 certificate and private key | TLS, mutual TLS, code signing |
| SSH host or user key identity | Identity asserted by possession of an SSH key | SSH key pair or SSH certificate | Servers, network devices, automation |
| Bot or automation account | Account used by a script, RPA robot or integration | Password, token | RPA tools, chat platforms, ITSM |
| AI agent identity | Identity given to an AI agent, ideally distinct from its user and its developer | Token, certificate or delegated credential | Agent platforms, identity providers, gateways |
Edge cases, decided
- Is an API key a non-human identity? No. It is a credential. The identity is the consumer the key was issued to. Where a platform issues keys with no consumer record, the key is the only trace of an identity that was never recorded, and that is a finding.
- Is a personal access token (PAT), or an API key issued to a person, a non-human identity? No. It is a credential of a human user account. When automation uses it, a human credential is being used non-interactively. Keep it in the NHI inventory as a credential to find and replace with an app or installation token or a federated identity, but do not count it as an NHI.
- Is a certificate a non-human identity? No. It is a credential that binds a key to a subject. The subject is the identity.
- Is an AI model an identity? No. A model produces outputs. An agent built on a model acts, and the agent needs an identity.
- Is a shared mailbox or break-glass account an NHI? It is a non-personal account used by people. Treat it under privileged access rules for humans, and keep it out of NHI counts.
- Is a device an NHI? A device can hold an identity, usually a certificate. Device management is a separate discipline; we include device identities only where a device calls services as a workload does.
- When an agent acts for a user, whose identity is it? Both should be visible: the agent's own identity and the user it acts for. If only the user appears in the log, the agent is invisible. If only the agent appears, the user is.
NHI and machine identity
"Machine identity" is the older term. In vendor usage it often centres on certificates and keys, though some vendors use it more broadly: CyberArk describes machine identities as spanning certificates, keys, secrets, workloads and more. "Non-human identity" is broader and includes accounts, roles, OAuth applications and agents. We use NHI as the umbrella term and say which type we mean. See NHI and machine identity compared.
How this taxonomy is used on this site
Company profiles list the identity and credential types each company says it covers, in each company's own words. Those lists are not yet mapped to the eleven types above. The risk model and the governance canvas use the same terms. The glossary holds shorter definitions of more terms.
Frequently asked questions
Is a secret or an API key a non-human identity?
No. A secret, such as an API key, token, password or private key, is a credential: the material an actor presents to prove an identity. The identity is the service account, OAuth client, workload or agent the credential was issued to. Governance attaches to the identity; rotation attaches to the credential.
Which is an example of a non-human identity: a service account, an employee login, a guest account or a visitor badge?
A service account. It is an identity created for software, not a person. An employee login and a guest account are identities assigned to people, and a visitor badge is a physical access token for a person. A service account with administrative rights is a privileged non-human identity.
Is a certificate a non-human identity?
No. A certificate is a credential that binds a key to a subject. The subject named in the certificate, such as a host, service or device, is the identity. The private key is the secret that proves it.
What types of non-human identity are there?
This taxonomy lists eleven: service accounts, service principals or application registrations, cloud roles or managed identities, workload identities, OAuth clients, API consumers, CI/CD identities, certificate subjects, SSH key identities, bot or automation accounts, and AI agent identities.
From the publisher
HumanAudit, which publishes this site, is preparing the NHI Governance pack: an NHI policy, ownership register, lifecycle procedures and one control set mapped, as our interpretation, to PCI DSS v4.0.1, ISO/IEC 27001:2022, NIS2 and DORA. Planned for March 2027; not yet on sale, and no price has been published.