What is a non-human identity?
A working definition, a boundary that survives contact with an auditor, and an honest map of how this overlaps with the identity categories you already own.
Published 2026-09-07 · Maintained by HumanAudit · Last reviewed 2026-09-07
The definition
A non-human identity is any credentialed principal that can authenticate to a system without a human present at the moment of authentication.
That boundary is chosen deliberately, because it is the exact point at which your existing controls stop functioning. Multi-factor authentication assumes someone can respond to a prompt. Session timeouts assume a session. Joiner-mover-leaver assumes employment. The quarterly access review assumes a manager who recognises the name in the row.
None of those hold for a service account. Which is why the population grows monotonically and nothing in the stack ever removes one.
What qualifies, and what does not
| Principal | Counts? | Why the boundary sits here |
|---|---|---|
| Service accounts and service principals | Yes | No person leaves the company, so no offboarding trigger ever fires. |
| API keys and static tokens | Yes | Bearer credentials. Possession is authorisation, and possession leaves no trace of intent. |
| OAuth applications and their grants | Yes | Consent is given once by one person; the scope persists indefinitely and survives their departure. |
| Certificates and SSH keys | Yes | Expiry is the only lifecycle event, and it arrives as an outage rather than a review. |
| Cloud workload identities, Kubernetes service accounts | Yes | Created by infrastructure-as-code, reviewed by nobody, and ephemeral enough that the population is a rate rather than a number. |
| CI/CD runners and pipeline identities | Yes | Hold the union of every secret the pipeline touches, which makes them the highest-value target in most estates. |
| AI agent identities | Yes | Act autonomously, at machine speed, on authority delegated by a human who is not present when they act. |
| Device and IoT identities | Yes | Enrolled in bulk, deprovisioned rarely. |
| A person's SSO account, even when scripted | No | A human is the accountable party and the normal joiner-mover-leaver controls apply. |
| A chatbot with no tool access and no credentials | No | It cannot authenticate to anything, so there is nothing to govern. |
The genuinely hard case
A human's personal access token used by automation — a GitHub PAT wired into a pipeline, a personal API key in a script. It has a human owner on paper and no human present in practice. Treat these as non-human identities for governance purposes: the failure mode is identical, and they are disproportionately represented in real incidents.
How this overlaps with what you already own
Ordered by what each category is actually about, rather than what its marketing claims.
| Category | Its real question | Covers NHI? | Where it stops |
|---|---|---|---|
| IAM | Who is this person and what may they access? | Incidentally | Built entirely around a human lifecycle. |
| PAM | How do we control the most dangerous accounts? | Yes, for privileged service accounts | Vaulting and session recording assume an interactive session. |
| Secrets management | Where does the credential live? | The credential, not the identity | Knows the secret exists; not who owns it or what it reaches. |
| Workload identity (SPIFFE, WIMSE) | Which process is this, cryptographically? | Yes, and correctly | Authentication only. Silent on authorisation and on intent. |
| CIEM | What effective permissions exist in cloud? | Cloud NHIs only | Absorbed into CNAPP by 2023. Blind to SaaS-to-SaaS OAuth. |
| ITDR | Is an identity behaving abnormally right now? | Increasingly | Detection, not governance or evidence. |
| IGA | Can we prove access was reviewed and approved? | Now, partially | Campaigns assume a human reviewer, a quarterly cadence, and attest to access rather than ownership. |
| API security | Is this API call malicious? | Traffic, not principals | Sees the request; not the identity's lifecycle. |
Common questions
What is a non-human identity?
A non-human identity, often abbreviated NHI, is any credentialed principal that can authenticate to a system without a human being present at the moment of authentication. Service accounts, API keys, OAuth applications, certificates, SSH keys, cloud workload identities, Kubernetes service accounts, CI/CD pipeline identities and AI agent identities are all non-human identities.
What is the difference between a non-human identity and a machine identity?
In practice the terms are used interchangeably. “Machine identity” is the older term and is preferred by certificate and PKI vendors; “non-human identity” emerged from startups around 2023 and is broader, since it comfortably includes OAuth applications and AI agents that are not machines in any physical sense.
Worth knowing: Gartner has not ratified the term “non-human identity”. Its July 2026 Hype Cycle for Digital Identity splits the same territory into Workload Identity Management, Workload Access Management and AI Agent Identity. KuppingerCole did publish a Leadership Compass for Non-Human Identity Management in November 2025.
Why do organisations have so many non-human identities?
Eight shifts multiplied the population rather than adding to it. Cloud made identity the perimeter, so every resource became something you authenticate to. SaaS moved the trust anchor outside your estate. Microservices turned internal function calls into authenticated network calls. Kubernetes gave every pod an identity and made pods ephemeral. CI/CD created build identities holding every secret a pipeline touches. Infrastructure-as-code made creating an identity a line in a file nobody reviews. Multi-cloud meant no single directory could see the whole population. And AI agents made the identity act on its own initiative.
We deliberately do not quote a machine-to-human ratio. Six vendors published figures between 17:1 and 144:1 within the same twelve months, every one self-published by a company selling machine-identity software, and we could locate no independent measurement.
Why does existing IAM not solve this?
Identity and access management is built around a human employment lifecycle: hire, move, leave. Those three events drive provisioning, review and deprovisioning. A service account has none of them. It has no manager to sign its access review and no leaving date to trigger its removal, so it persists by default until somebody deliberately hunts it down.
What is an AI agent identity?
An AI agent identity is the credential an autonomous or semi-autonomous AI system uses to act on other systems — reading a mailbox, writing to a repository, querying a database, calling a tool through a protocol such as MCP. It is a non-human identity with two properties that make it harder to govern: it acts without a human present, and it acts on authority delegated by a human whose intent is not encoded anywhere in the credential.
This distinction matters. In every documented AI-agent security incident of 2025 and 2026, the credential was correctly issued, correctly scoped and correctly authenticated. What failed was not identity but accountability.
What is secrets sprawl and how does it relate?
Secrets sprawl describes credentials proliferating across code, configuration, CI/CD systems, vaults and chat. It is the credential-side view of the same problem. GitGuardian, which measures it directly, found 28.65 million new hardcoded secrets added to public GitHub repositories during 2025, and that 64% of valid secrets first leaked in 2022 were still active in January 2026.
Secrets sprawl tells you a credential exists and is exposed. It does not tell you which identity it belongs to, what that identity can reach, or who is accountable for it. Those are the questions on the other side of the same problem.
How do you govern non-human identities?
In practice, in four stages. Discover every non-human principal across your identity providers and SaaS admin surfaces. Attribute each one to an accountable human, and record explicitly where none exists. Assess scope against actual use, since permissions never exercised are the cheapest to remove. And retain the record, so that ownership at a past date is a question with an answer.
Rotation, least privilege and short-lived credentials are all worth doing and all downstream of the first stage. You cannot rotate what you cannot enumerate, and you cannot assign an owner to something you have not found.
Find out how many of yours have no owner
A read-only guided scan across Google Workspace, Microsoft 365, GitHub and Slack. You keep the report regardless of what happens next.