A Non-Human Identity Security Manager Is Hired to Delete Credentials Nobody Will Claim

The Service Account That Deployed to Production Last Night Has No Owner

At 2am a service account pushed a change to production. It was created in 2021, its name is a six-character string, it holds write access to three data stores, and the engineer who made it left in 2023. Nobody can tell you whether turning it off breaks billing. That single unanswerable question, repeated a few thousand times, is the condition a Non-Human Identity Security Manager is hired into.

Screen first for a bias toward discovery over declaration. Ask a candidate how they would build the inventory. A weak answer describes a request form and a registration policy, which produces a list of the identities people remembered to tell you about. A strong answer starts from the systems themselves: cloud audit logs, identity provider records, certificate transparency, secret manager contents, CI runners, and then reconciles what is actually authenticating against what anyone claims exists. The gap between those two sets is the real finding, and a candidate who has done this once will volunteer that the gap was embarrassing.

Comfort with deletion comes next, tempered by blast-radius reasoning. The tell is a specific story: which credential they revoked, what they checked first, what broke anyway, and how long recovery took. Listen for the ninety-day usage window they pulled before deciding, the owner they paged, and the rollback path they kept open for a week. An answer built out of governance frameworks belongs to someone who has never turned anything off.

Negotiation is the part most searches skip, and this role cannot ship anything alone. Every fix lands in a team that did not ask for it. Watch how a candidate describes getting a platform team to rotate a shared key it had been reusing for years. A policy memo and an escalation predicts a stalled program. An offer to do the rotation themselves in a staged window is the person you want. The same instinct is what an MLSecOps engineer is hired for, rather than the filing of findings.

Analysts are treating this as a defining 2026 problem rather than a housekeeping backlog, with agentic AI and the non-human identity surge named together as what security teams and their providers will be wrestling with ¹. The arithmetic is what makes it urgent: in cloud-native environments non-human identities outnumber human ones by at least fifty to one, and the trend reporting puts the real number higher than that ².

Look at Certificate Operations Before Any Security Title

The obvious feeders work: IAM engineers, PKI and certificate operations, secrets management owners, cloud platform and SRE staff who wrote the workload identity federation nobody else understands. What all four share is that they have already been on the hook when an expiry landed badly, which is the fastest available teacher for how machine credentials fail.

Certificate operations converts best and gets passed over most often, because the title sounds like a maintenance job. Someone who ran a fleet through a certificate expiry outage understands the whole shape of this work already: an inventory that was wrong, an owner who had moved teams, an automated renewal that silently stopped, and a business impact that arrived without warning. Machine identity is that problem with more identity types and a faster creation rate.

The unexpected backgrounds are worth chasing. Telecom and device provisioning engineers have managed millions of credentialed non-people for decades and bring lifecycle instincts nobody in a cloud team was taught. IT asset management staff have already fought the exact fight this role fights, which is getting an accurate inventory of things that appear without paperwork. Operational technology engineers arrive knowing that you cannot simply revoke something connected to a physical process, which is the judgment that keeps a clean-up from causing an incident. Payment operations people who have sat through a key ceremony understand custody in a way that transfers directly.

What almost nobody arrives with is the agent case. A service account is created once and behaves the same way for years. An agent holds delegated authority, chains calls across systems, and may request scope at runtime, so the question shifts from who owns this credential to what was this identity allowed to decide on someone's behalf. That is closer to the delegation problems an AI system auditor works on than to classic account hygiene, and the governance vacuum around it is documented: analysis of non-human identity and agentic AI found a meaningful share of organizations, more than sixteen percent, not tracking AI-related identity creation at all ³.

Ask How They Used AI Against Their Own Credential Sprawl

Ask what they did with an assistant on this exact problem, and listen for a workflow rather than a stance. The strong answers are unglamorous: pulling ninety days of audit events and having a model cluster them into which identities actually called which actions, then generating a least-privilege policy diff from the observed calls, then reading every line of that diff before anyone applied it.

The tell that matters is what they check. A model will confidently produce a permission name that does not exist, or a policy that parses cleanly and grants more than the previous one. A candidate who has done this at scale will say so without being prompted, and will describe the check they run every time: validate the generated policy against the provider's own evaluator, then replay real traffic against it before it goes near production. That habit of testing a claim against something outside the conversation is the difference between an assistant that saves a week and one that writes an incident.

Press on the reverse direction too, because agents are now the thing being governed as well as the tool. Ask what they would require before granting an agent framework its own credential. Good answers name a human owner, a scope derived from a specific task rather than a role, a short expiry, and a log that ties every action back to the person whose authority was delegated. Weak answers give the agent a role that already exists, which is how a coding assistant ends up holding the same access as the platform team that runs it.

One caution about interviewing this. The vocabulary here is easy to hold and hard to distinguish from experience. Workload identity federation, secret zero, ephemeral credentials, intent-based access: a candidate saying all four may have run a program or read a vendor page last week. Hand them a real export from one of your accounts with the names removed, give them ninety minutes, and ask which twenty identities they would kill first and what they would check before each one.

Find NHI Managers Where Credentials Get Rotated, Not Where Titles Get Posted

Start inside. The person who already owns your secret manager, your certificate authority, or your cloud landing zone has most of this job's context and usually wants the mandate. Second internal pool: whoever ran your last access review and can tell you which parts of it were theater.

Outside, the professional venues are narrow and real. IDPro is the practitioner body for identity work and its members skew toward people who have actually operated these systems. The Cloud Security Alliance publishes the research this discipline argues about. The SPIFFE and SPIRE community under the CNCF is where workload identity gets debated by people implementing it, and the OAuth working group at the IETF is where the delegation standards agents will use are being written. Regional BSides conferences surface operators who do not speak at the large vendor events. Adjacent titles worth approaching directly: IAM engineer, PKI lead, cloud security architect, platform security engineer, secrets platform owner.

What they care about is authority, and the offer usually dies on that point rather than on money. A role that owns the inventory but cannot revoke anything is the one every experienced candidate recognizes in the first interview and treats as a career trap. Reporting into the team whose delivery dates the role will block kills it just as reliably. So does being handed a purchased tool and told the program is the rollout, when the actual work is ownership assignment across dozens of teams that have never met you.

What closes it is concrete. Name the standing forum where an unowned identity gets a decision and who chairs it. State in writing that the role can revoke after a defined notice window, and name the exception path. Give a first-year target that is a reduction rather than a report: orphaned credentials eliminated, average credential lifetime, share of identities with a named owner. The 2am account should be gone or owned by month four, and that is a promise somebody can check. And be honest that the first six months are inventory and unpopular conversations, because the candidates who last are the ones who knew that going in. Where agents are being bought rather than built, this person ends up in the room with your procurement agent orchestrator deciding what a vendor's agent is allowed to hold.

What Does Revocation Authority Cost You?

Revocation authority is the thing you are actually buying, and it is the reason no published band fits. Nothing on the market prices a security manager by what they are allowed to switch off. So build the number from your senior security engineering or IAM lead band, because that is the pool this candidate is leaving, then add for the operational risk and be ready to say what the addition is for.

That addition has a defensible shape. Someone who can disable a credential that stops a revenue system carries the same class of risk you already pay for in on-call platform leadership, and pretending otherwise is how the seat gets filled by a person who will never use the authority. If the hire will also write the discovery pipeline and the policy generation rather than governing someone else's, you are bidding against platform engineering offers and should price closer to an AI platform engineer than to a compliance band.

Do not read a current title as evidence of any of this. Two people carrying the same machine-identity title may have had opposite mandates, one keeping a spreadsheet and the other holding a kill switch. What separates them is what they switched off last year and who they had to tell first.

On location, the work is remote-friendly by nature: audit logs, policy code, and reviews with teams who are themselves distributed. Two things resist it. Incident response during a botched rotation goes better with overlapping hours, so buy timezone overlap rather than a desk. And key custody in regulated environments sometimes requires physical presence, because a hardware security module ceremony genuinely happens in a room. Scope that before writing the offer, since it decides which candidates can take the job at all.

On-premise pressure comes from the same place. Where credentials protect regulated data or industrial systems, the tooling has to run inside the boundary and the person may need site access on a schedule. One flag rather than advice: retention, audit-trail and access-review duties for automated systems differ by jurisdiction and are still moving through 2026, and the records this role produces are frequently the only evidence of who authorized what an agent did. Check with counsel in your jurisdiction rather than reasoning from a summary.

Read the evidence >