--- title: "Non-Human Identity Security: Score Your Agent Credential Lifecycle" description: "Eight questions on owners, end dates and offboarding, scored in an hour. Plus the 2026 scanning data that shows why expiry beats discovery as a first purchase." author: "Renn Calloway" category: "Defensive Architecture & Security Controls" date: 2026-10-24T13:15:00.634Z canonical: "https://agenticcyber.co/blog/non-human-identity-lifecycle-for-agents-owners-expiry-and-offboarding" --- # Non-Human Identity Security: Score Your Agent Credential Lifecycle ![agent orange victim, people with disabilities, sos peace](https://cdn.pixabay.com/photo/2022/12/29/01/35/agent-orange-victim-7684015_1280.jpg?w=1200&q=75) > Eight questions on owners, end dates and offboarding, scored in an hour. Plus the 2026 scanning data that shows why expiry beats discovery as a first purchase. One sentence closes more identity reviews than any other: we scoped the permissions, so that one is covered. Scope is half of it. The skipped half is duration: how long the thing goes on existing, which standing team is accountable, and where it ends up once the project behind it is cancelled. What follows scores the duration half. Eight questions, nought to two points each, and a two only when somebody in the room can produce the evidence. Products exist that do the discovery part of this at scale, and some teams will want one. This costs an hour and tells you whether you need one yet. ## The Eight Questions, Two Points Each #Question012, with evidence 1Does one register hold every machine credential an agent uses: service accounts, keys, OAuth grants, certificates?Nothing existsOne or two systems coveredA single export carrying the date each credential was minted and the date it last authenticated 2Does each entry name a standing team, plus a way to reach it outside office hours?No ownership fieldAn individual's name sits in the fieldA team owns it, with a documented escalation path 3Could you produce, today, every credential that has not authenticated in a quarter?Not knowableKnowable for part of the estateA scheduled last-authentication report with a named reader 4Does the identity itself carry an end date, separate from how long its tokens live?Identities do not lapseSome lapseEvery identity lapses unless its owning team renews it 5When an agent is retired, does a written step cover removing its identity?No procedureA procedure exists and omits the identityRevoke grants, destroy keys, clear every registry, then verify, all written down 6Are the scopes each credential holds today compared against what its agent needs today?NeverOnce, at launchA recurring drift review, with every removal recorded 7Are credentials whose expiry is distant or absent enumerated and on a retirement schedule?Not enumeratedEnumerated, nothing retiredEnumerated, each with a retirement date against it 8Have you destroyed a test agent and then gone looking for what survived it?NeverOnce, checked partiallyWithin two quarters, findings written up ## Reading the Total - **Nought to five.** Stale and ownerless credentials almost certainly sit in your estate right now. The problem is less that they exist and more that you cannot demonstrate otherwise. - **Six to eleven.** You provision well and you forget well. Sizing happens at launch and then nothing happens at all. - **Twelve to sixteen.** Every credential now carries a holder, a fixed term and an ending. The remaining work is keeping the drill honest. Those bands are a working heuristic from this desk, not a threshold anybody has published. ## Where Each Band Starts ### Nought to five - Build the register. Owner, purpose, date minted, date last authenticated. That final column surfaces the dead ones faster than any other method available to you. - Move ownership onto teams. A credential that never belonged to an individual cannot be orphaned by a resignation. - Publish a claim deadline. Whatever is unclaimed when it passes gets disabled, then deleted. ### Six to eleven - Put an end date on the identity, renewable only against a positive confirmation from the owning team that the agent is still running. - Write identity removal into the decommission procedure, beside tearing down the infrastructure, and make the last line a verification rather than an assumption. - Run a first drift review: what the credential can reach now, against what its agent needs now. The permission half is covered in our [roadmap from role-based towards attribute-based access](/from-rbac-to-abac-a-practical-access-control-roadmap-for-ai-agents-and-autonomou). ### Twelve to sixteen - Destroy a test agent and document everything that outlived it. - Put the distant-expiry and no-expiry credentials on a retirement calendar, and keep that calendar where somebody actually looks. - Fold the register into the inventory you already maintain, using the approach in [mapping your surface before anybody else maps it](/how-to-map-your-attack-surface-before-attackers-do), pointed at credentials instead of hosts. ## The Evidence Behind the Questions Five of the eight come out of the OWASP list for non-human identities, whose top entry concerns removal rather than scope. [NHI1:2025 addresses machine identities left in place, neither deactivated nor removed](https://owasp.org/www-project-non-human-identities-top-10/2025/1-improper-offboarding), once nobody needs them, and it names two shapes: a credential that outlives the application it was minted for, and one abandoned when its owner moves on. Questions one, two, three and five probe both. Agents generate these faster than anything before them, because starting an experiment is free and nobody books its end. Question six follows [NHI5:2025, where OWASP observes that overprivileged non-human identities skew towards administrative accounts](https://owasp.org/www-project-non-human-identities-top-10/2025/5-overprivileged-nhi), with the impact that implies when one gets abused. Old, forgotten and wide is the combination worth hunting first. Question seven follows [NHI7:2025 on long-lived secrets](https://owasp.org/www-project-non-human-identities-top-10/2025/7-long-lived-secrets), covering keys, tokens and certificates whose expiry sits implausibly far out or simply does not exist, and noting that a leaked one grants access ["without any time constraints"](https://owasp.org/www-project-non-human-identities-top-10/2025/7-long-lived-secrets). Questions two, four and eight are judgement rather than citation. A short token lifetime bounds how long a theft stays useful and does nothing whatever to the identity behind the token. An end date on the identity does, and a drill is the only thing that proves the end date fires. The runtime half of this sits in [our piece on agents and machine credentials](/non-human-identity-security-for-ai-agents), and keeping the secret out of the configuration to begin with is [its own discipline](/secrets-management-for-ai-agents-keep-credentials-out-of-configs). ## Most Leaked Credentials Were Never Revoked Question seven carries more weight than two points suggests, and the current measurement explains why. GitGuardian's fifth annual secrets report, published in March 2026, [counted 28.65 million new hardcoded secrets in public GitHub commits across 2025, a 34 per cent rise year on year](https://blog.gitguardian.com/the-state-of-secrets-sprawl-2026/). That is the volume story, and it is the one that gets quoted. The duration story is the one that belongs next to your scorecard. In the same report, [nearly 70 per cent of the credentials GitGuardian had confirmed valid back in 2022 were still valid when retested in January 2025, and the number remained above 64 per cent when they retested a year later](https://blog.gitguardian.com/the-state-of-secrets-sprawl-2026/). Read that as a revocation rate arrived at by subtraction. Most leaked credentials are never withdrawn, and a meaningful share of what leaked four years ago is still authenticating today. That is precisely what an estate with no identity-level expiry looks like from the outside. Two further figures from that report should change what you examine first. Internal repositories are roughly six times more likely than public ones to hold a hardcoded secret, [on GitGuardian's own scanning](https://blog.gitguardian.com/the-state-of-secrets-sprawl-2026/), so an estate that feels clean because its public code is clean has not been measured at all. And secrets tied to AI services climbed 81 per cent year on year, to just over 1.27 million during 2025, [per the same report](https://blog.gitguardian.com/the-state-of-secrets-sprawl-2026/), which is the agent-shaped portion of this arriving in the data. On process, Okta's lifecycle guide, updated in February 2026, is worth reading for one pattern in particular. It describes suspending any credential that has not authenticated inside a defined window, offering thirty days as the example, and then [holding an identity in quarantine for twenty-four hours before permanent deletion so that dependencies surface before the delete lands](https://www.okta.com/identity-101/what-is-the-non-human-identity-lifecycle/). That quarantine step is what most home-grown offboarding lacks, and its absence is why teams stop at disable and never reach delete. The same guide notes that identities declared in infrastructure code can be revoked when the resource they served is destroyed, which answers question five by construction rather than by procedure. ## Discovery Tools Find the Problem. Expiry Removes It. Here is the call, and it cuts against how this category normally gets bought. A team scoring badly on this check will be sold discovery first, because discovery demonstrates beautifully: a dashboard appears, the count is shocking, everybody agrees that something must be done. Fund question four ahead of it. An identity that lapses unless its owning team renews it turns the whole thing from a detection problem into a default, and the default keeps working through the quarter when nobody opens the dashboard. The argument for that ordering is sitting in the validity figures above. Discovery has been purchasable for years, and credentials from 2022 are still live, [by GitGuardian's own retest](https://blog.gitguardian.com/the-state-of-secrets-sprawl-2026/). Finding them was never the hard part. What nobody had was an owner obliged to renew and a clock that ran out when no owner did. Add the clock and the register shrinks without anybody being chased. A specific situation undoes this. If you genuinely cannot enumerate your credentials, expiry is unbuildable, because nothing can expire what nobody can list. Scoring a zero on question one means buying or building the discovery first and coming back to this. The same is true of an estate where every credential is already minted by infrastructure code with a short attested lifetime. The clock exists there already, and discovery is the only gap left. ## Who Should Be Holding the Scorecard Scoring alone produces a flattering number. Three roles need to be present for half an hour: the operator of your identity provider, somebody who owns a live agent, and an engineer from security. Work down the table out loud, and require the evidence before anybody states a number. Disagreement is the useful output, because two people scoring question two differently have located drift between them. Write the total and the date where all three can find it, repeat the exercise whenever an agent is retired, and watch the trend. A rising total matters less than a falling count of credentials nobody will claim. ## The Limits of a Number This measures process and not exposure. An owned credential with an end date against it can still hold far more access than its term would justify. Nothing here reveals whether a particular secret has already leaked, which is what scanning exists to answer. This is no audit, it rates no product, and decisions about production risk belong with qualified security professionals who can see the environment in front of them. ## Questions about leasing an agent identity ### How is identity expiry different from token expiry? A token lifetime bounds the useful window after a theft. The identity standing behind that token carries on regardless. Identity expiry ends the thing itself unless its owner renews, so an agent nobody remembers removes itself. ### Why put a team in the owner field rather than a person? Individuals change teams and leave companies. A standing team with an escalation path cannot resign, so the credential never becomes ownerless in the first place. ### How do dormant agent credentials get found? Record a last-authentication date against every credential and review that column on a schedule. Dormancy shows up there before it shows up anywhere else. ### Does this check replace a non-human identity product? No. It tells you where you stand and what to repair first. Discovery at scale is a reasonable purchase, once the expiry default exists. ## FAQ ### How is identity expiry different from token expiry? A token lifetime bounds the useful window after a theft. The identity standing behind that token carries on regardless. Identity expiry ends the thing itself unless its owner renews, so an agent nobody remembers removes itself. ### Why put a team in the owner field rather than a person? Individuals change teams and leave companies. A standing team with an escalation path cannot resign, so the credential never becomes ownerless in the first place. ### How do dormant agent credentials get found? Record a last-authentication date against every credential and review that column on a schedule. Dormancy shows up there before it shows up anywhere else. ### Does this check replace a non-human identity product? No. It tells you where you stand and what to repair first. Discovery at scale is a reasonable purchase, once the expiry default exists. --- Source: https://agenticcyber.co/blog/non-human-identity-lifecycle-for-agents-owners-expiry-and-offboarding