Sourced guideLast checked 16 October 2026 with AI-assisted research. Editor: Akshay Dubey.Corrections logEditorial policy
On this page
  1. Why SOC 2 is different
  2. The CC6 criteria
  3. The period problem
  4. Where exceptions come from
  5. Preparing for a Type II
  6. The population problem
  7. Type I against Type II, control by control
  8. FAQ

TL;DR

  • A Type II report tests operation across a period, typically six to twelve months. Fixing a control the week before fieldwork does not help; it may make the exception more visible.
  • The relevant criteria are the CC6 series on logical access, and CC6.2 and CC6.3 are where non-human identities land hardest.
  • Evidence retention is the binding constraint. If your logs roll off at 90 days and the period is 12 months, you cannot evidence the first nine.
  • An exception to expect is terminated-owner service accounts, because auditors sample leavers and trace their access.
  • SOC 2 is an attestation, not a certification. There is no pass mark; there are exceptions, and how you describe them matters.

Why SOC 2 behaves differently from a certification

An ISO certification audit largely asks whether a management system exists and functions. A SOC 2 Type II asks an auditor to test samples of control operation throughout a defined period and to report what they found, including what they found wrong.

For non-human identity this changes the work in three ways. Remediation has a lead time, a control implemented in month ten produces two months of evidence in a twelve-month period, and the report will say so. Evidence must survive, so retention policy becomes a control dependency. And exceptions are published rather than resolved privately, which means the report your customers read may name the gap.

The CC6 criteria, read for machine identities

Our practical reading. The Trust Services Criteria are written generically and do not name non-human identities; auditors apply them to whatever accesses the system. The second column is our own short summary. The criteria are AICPA copyright and we do not reproduce their text; read the criteria themselves for the wording your auditor works from.

SOC 2 common criteria for logical access and monitoring, read for non-human identities
CriterionIn shortWhat gets sampled for NHIEvidence
CC6.1Logical access security over protected assets.Whether everything that accesses the system is known, including service accounts, integrations and workloads, and whether machine access is restricted and authenticated. Where credentials are stored, and whether any exist outside a controlled store.An inventory of what accesses the system, not only user accounts. Configuration evidence for authentication mechanisms. Secrets store coverage against the full identity population. Secret scanning results across repositories, pipelines and images.
CC6.2Registration and authorisation before credentials are issued.Approval before creation. For a sample of non-human identities created during the period: was there an authorisation, and did it come before issuance? This is where ad hoc creation becomes visible.Creation records with approver and timestamp, tied to the identity. Reviewed infrastructure-as-code merges count, and are usually stronger evidence than tickets.
CC6.3Access is changed and removed when it should be.Review and removal. Was the periodic review population complete, and were identities removed when the workload they served was retired or their owner left? This is the hardest criterion to evidence after the fact.Review records naming the reviewer and the decision per identity, removal records with timestamps, and dormancy reports showing no authentication after removal.
CC6.6Protection against threats from outside the system boundary.Which non-human identities can be reached or used from outside, and how strongly they authenticate.A list of externally reachable machine credentials with their authentication mechanism and any network restriction.
CC6.7Restricted transmission and movement of information.Which integrations move data out of the system, under what grant, and how often that grant is reviewed.A register of outbound integrations and OAuth grants with scope, owner and last review date.
CC6.8Prevention and detection of unauthorised software.Controls over what runs with credentials, including build pipelines.Pipeline configuration showing which jobs can obtain which credentials, and change approval for it. See CI/CD credentials.
CC7.2Monitoring for anomalies.Whether detection covers machine credential use and not only user sign-ins.Detection rules for non-human identities and a sample of alerts raised and handled during the period.

Scope note

Which criteria apply depends on the trust services categories in your report. Security (the common criteria) is always included; Availability, Confidentiality, Processing Integrity and Privacy are elective. The mapping above assumes Security. If your report includes Confidentiality, expect additional scrutiny of which non-human identities can reach confidential data and how that is constrained.

The period problem

This is the failure most specific to SOC 2 and the least anticipated.

  • Log retention shorter than the period. If cloud audit logs roll off at 90 days and the observation period is 12 months, the evidence for months one to nine does not exist. No amount of current-state excellence recovers it. Check retention against period length before the period starts.
  • Controls implemented mid-period produce partial evidence and the report describes the timing. This is survivable and it is better to state the implementation date than to have the auditor discover it.
  • Population completeness. The auditor tests a sample from a population you provide. If your population of non-human identities is incomplete, the sample is drawn from an incomplete list, and a later discovery that the population was wrong is a materially worse finding than an exception within a complete one.

Where exceptions actually come from

  1. Terminated-owner service accounts. Auditors sample leavers and trace access. Machine identities the leaver created or owned are the classic hit, and they are hard to remediate retroactively because the evidence of who owned them is what is missing.
  2. Credentials issued without an approval record under CC6.2. Common where engineers create service accounts as part of normal work and the approval exists only as a merged pull request nobody mapped to the criterion.
  3. Access reviews with no non-human population. Same root cause as the ISO finding, and it surfaces here as an incomplete population rather than a missing control.
  4. Non-expiring credentials with no documented exception. Not automatically an exception if registered and approved; reliably one if invisible.
  5. Third-party integration grants outside the register under CC6.7.

Preparing for a Type II with NHI in scope

  1. Set retention before the period begins. Longest single lever, and it costs money rather than effort, which makes it easy to get wrong late.
  2. Establish the population and freeze how it is generated. An auditor will ask how the list was produced and whether the method could miss identities.
  3. Make approval a system record, not a conversation. If credential creation runs through a request path, CC6.2 evidence is automatic. If it runs through engineers, you will be reconstructing it.
  4. Wire the leaver process to the identity registry. This change removes that category of exception.
  5. Run one review with the non-human population included early in the period, so the pattern is established rather than introduced late.
  6. Register exceptions with an approver and an expiry. A documented, approved deviation is a governance decision. An undocumented one is a finding.

SOC 2 does not require perfection. It requires that what you claim is what happened, evidenced across the period, with deviations disclosed. Non-human identity often fails that test, and usually for evidential rather than technical reasons.

The population problem

In our reading, many substantive machine-identity exceptions trace back to one question the auditor asks early: how do you know this listing is complete?

The usual answer is an export from the identity provider. For human users that is close to authoritative, because joiners flow through HR. For non-human identities it is a fraction of the population, because identities are minted in places the IdP never sees:

Scroll sideways to see all columns.

SourceTypically in the IdP export?How it gets created
Entra ID / Okta service principals, app registrationsYesAdmin or app registration flow
AWS IAM roles and users, Azure managed identities, GCP service accountsNoConsole, CLI or Terraform, by whoever owns the account
Kubernetes service accountsNoManifest apply; created by default in every namespace
Database and middleware accountsNoDBA or migration script
CI/CD system credentials and deploy keysNoRepository or pipeline settings
Third-party OAuth grants into your tenantPartlyUser consent, often without admin involvement

Why this compounds

An incomplete population does not produce one exception. It weakens every test performed on that population, because the auditor cannot conclude that a sample drawn from an incomplete listing is representative. A single completeness gap can turn three clean control tests into three qualified ones. This is the highest-leverage thing to fix, and it is fixable months before fieldwork.

Demonstrating completeness

What auditors accept, roughly in ascending order of strength:

  1. An attestation that the listing is complete. Weakest; relies on your assertion.
  2. An enumerated source list naming every identity store queried, with the query used. Better, because it is testable.
  3. Automated discovery across those sources on a defined schedule, with the run evidence retained. Stronger, and reusable every period.
  4. Reconciliation between discovered identities and authentication events observed in logs, with unexplained authenticating principals investigated. Strongest, because it detects identities your discovery missed.

See discovery and inventory for how to build this.

Type I against Type II, control by control

Type I asks whether the control is suitably designed at a point in time. Type II asks whether it operated effectively throughout the observation period. For machine identities the second question is materially harder, and for a specific reason: identities are created continuously, by automation, outside the process your control describes.

Scroll sideways to see all columns.

ControlType I can be satisfied byType II requires
Authorisation before creationA documented approval process and a current inventoryApproval evidence for a sample spanning the whole period, including identities created by pipelines at 3am
Periodic access reviewA defined review procedure with scope covering NHICompleted reviews at every defined interval in the period, on complete populations, with decisions implemented
Removal on decommissioningA defined removal procedureRemoval records for a sample of workloads retired during the period, with timestamps
Credential managementA rotation policy and a secrets storeRotation or issuance records demonstrating the policy operated, with exceptions tracked

The practical implication

If identities can be created outside the approved path, the honest options are to close the path or to change the control description. A control that says every identity is approved before creation, in an environment where Terraform can mint one without a ticket, will generate exceptions for the whole period. Redefining the control around the merge review that approves the IaC change is often both truer and more testable than pretending the ticket flow covers it.

Frequently asked questions

How does SOC 2 treat non-human identities?

The Trust Services Criteria are written generically and do not name them, so auditors apply the CC6 logical access criteria to whatever accesses the system, including service accounts, integrations and workloads. CC6.2 on registration and authorisation before issuing credentials, and CC6.3 on modifying and removing access, are where non-human identities land hardest, because both require records that most organisations generate only for human accounts.

What is the difference between SOC 2 Type I and Type II for machine identity?

Type I assesses control design at a point in time; Type II tests whether controls operated effectively across a period, typically six to twelve months. For non-human identity this is the decisive difference, because it converts the question from whether you have a control to whether you can evidence it operating throughout, which depends on log retention, population completeness and records generated at the time rather than reconstructed later.

What SOC 2 exception should we expect for non-human identities?

Service accounts owned or created by terminated employees. Auditors sample leavers and trace their access, and machine identities attached to those leavers are the classic hit. They are hard to remediate retroactively because the missing evidence is precisely the record of who owned them, which is why wiring the leaver process to the identity registry removes more exceptions than any other single change.

Why does log retention matter for SOC 2 and non-human identity?

Because a Type II report covers a period. If cloud audit logs roll off at 90 days and the observation period is 12 months, evidence for the first nine months does not exist and no amount of current-state excellence recovers it. Retention should be checked against the intended period length before the period starts, since it is one of the few levers that cannot be pulled retroactively.

Is SOC 2 a certification?

No, it is an attestation. An independent auditor reports on what they tested and what they found, including exceptions, and there is no pass mark. That is why how exceptions are described matters: the report your customers read may name the gap, and stating an implementation date plainly is generally better than having the auditor discover the timing.

Does SOC 2 cover service accounts and machine identities?

Yes. The logical access common criteria CC6.1 to CC6.3 concern authorising, registering, modifying and removing access to protected information assets, and nothing in them limits the subject to human users. In our reading, a service auditor can include service accounts, API credentials and cloud IAM principals in the population they sample.

What NHI evidence does an auditor ask for?

Population completeness first, then samples: a listing of all accounts with access to in-scope systems including non-human accounts, approval evidence for accounts created during the period, review evidence covering the same population, and removal evidence for accounts deprovisioned during the period.

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.

Page history

Merged from the nhimanagement.com guide on 7 October 2026, when the two sites were combined: the sections "The population problem"; "Type I against Type II, control by control" and 2 questions in the FAQ. The merged material is under review against our methodology.

Monthly change log

Know when the NHI landscape changes.

One email a month from this reference: new deals, standards status changes and corrections. No sponsored content. Unsubscribe from any email.

We first email you a link to confirm. Privacy policy.