Skip to main content
INS // Insights

Least Privilege Baseline Review

Updated October 2026 · 7 min read

A least privilege baseline review compares the grants an identity holds with the template for that job. The question is not "does this look reasonable." The question is "which seats are outside the catalog we already agreed." A reviewer who answers from memory will approve the extra admin group because the person is senior. The template does not care that they are senior. If the function is support reader, production admin is drift, and the row says so before anyone clicks keep.

Rutagon builds the catalog and the diff. The GRC platform the team already runs stores the decision. It does not know the template unless we load it. Security automation is the diff. Privilege creep access review is what the extra seats become over time. The baseline review is the comparison that makes the creep visible against a versioned list, instead of against last quarter's approvals.

What a least privilege baseline review compares

The catalog is a function key and the grants that function should have. A function is a job shape, not a person: support reader, on-call, billing, deploy for one service. Each function maps to permission sets, groups, or role names. The catalog also names what is forbidden for that function, and whether a permission boundary is required. A person is in the review only after we know their function. A person with no function is an unowned row. They are not squeezed into the nearest template so the campaign can finish.

Drift is a set of differences, not a vibe. An extra group or permission set that the template does not name. A missing grant the template says is birthright, which is a joiner problem more than a privilege problem, and we still show it so the template is not fiction. A direct policy on an IAM user when the template says grants come only from groups. A wildcard action or resource that the template's policy does not contain. A missing permission boundary where the template requires one. A local exception whose end date has passed.

The template is code. A pull request changes it. A person's grants changing outside the template is drift, even when the new grant would have been a sensible template edit. We do not rewrite the template to match production so the diff goes green. That is how the baseline becomes a photograph of privilege creep. The review can propose a template change. Until that change merges, the extra seat stays a finding.

IAM Identity Center access review is where permission set assignments are listed. The baseline uses those assignments as the actual side of the diff when Identity Center is the grant path. An assignment in an account that is not in the catalog, including a one-off permission set someone created for a weekend, is drift. Console grants and Identity Center grants are both actuals. Reading only one of them hides the other.

function: support-reader
template:
  - permission-set/prod-readonly
  boundary_required: true
actual:
  - permission-set/prod-readonly
  - permission-set/prod-admin
boundary_present: false
drift:
  - extra permission-set/prod-admin
  - missing boundary
decision: remove extra, attach boundary

A permission boundary does not grant anything. AWS describes it in Permissions boundaries for IAM entities. It is a ceiling. The identity policy and the boundary both have to allow an action. The baseline requires a boundary on roles a developer can create or pass, because the next policy version can widen up to that ceiling. With no boundary, the next version can widen to full administration. A role whose current policy looks narrow and whose boundary is absent is still drift. The current policy is not the control. The ceiling is.

Effective access is not the same as attached access

A service control policy can deny what an identity policy allows. SCP least privilege is that ceiling at the organization. The baseline review still flags the attach. An admin permission set that an SCP denies is not "fine." It is a grant that will start working the day the SCP changes, and it is a grant the reviewer can see in the account. The row has two columns: attached, and effective. Effective is what a simulation or the SCP intersection says will work today. The decision is about attached. We do not auto-maintain a forbidden attach because it is currently ineffective.

The same split applies to a permission boundary that already caps a wide policy. The wide policy is drift against a template that does not contain those actions. "The boundary saves us" is a reason to keep the boundary, not a reason to keep the wide policy. Both columns stay on the row so a reviewer cannot approve the policy by pointing at the boundary, or approve a missing boundary by pointing at today's narrow policy.

Workload roles are a separate catalog. A deploy role for a pipeline is not a job function a person should be assigned. On a production SaaS platform we operate, human functions and workload roles are different lists, and a person showing up with a deploy role is drift against the human template. The workload catalog is reviewed on its own cycle. Mixing them lets a pipeline role look like a job title.

Group membership is expanded before the diff. A template that names a group, and a person who is in that group through a nested group, is not automatically clean. We store the path. Nested membership the template did not name is drift even when the resulting group is one the template allows by a different path. The owner can accept the path. They accept a path, not an invisible one. Direct membership that duplicates a nested path is still listed. Duplicate paths are how removal fails later: someone removes the direct add and the nested group remains.

Decisions and evidence

The decision is per drift item, not per person. Remove the extra permission set. Attach the missing boundary. Keep, with a reason and an end date, when the template is wrong and the pull request is open. A keep with no template change and no end date is rejected. The read-back for a remove is the assignment gone. The read-back for a boundary is the boundary attached on the role. A ticket that describes the intent, with the assignment still present, stays open.

The evidence export is the template version, the function for each person, the drift items, the decider, and the read-back. People with no function are a separate table, not a footnote. An auditor who samples "show me least privilege for production admin" gets the function that is allowed to hold it, the people who hold it, and the people who hold it without that function. That third list is the review.

We do not generate the template by clustering current grants. Clustering photographs the creep and calls it a baseline. The catalog starts from the functions the company can name. Where a function cannot be named, the identity stays unowned until an owner names it. Unowned is a valid output of the first review. It is not a failure of the diff.

See what an Access & Credential Governance Diagnostic finds outside the role templates you think you use. Contact Rutagon at rutagon.com/contact, 907-841-8407, or contact@rutagon.com.

Frequently Asked Questions

Is this the same as a quarterly access review?

No. A quarterly review asks an owner to affirm each entitlement. A baseline review asks whether the entitlement is in the template for the function. A seat can be affirmed for years and still be drift. The baseline diff is the list. The quarterly decision is what happens to each drift item.

Does a permission boundary satisfy least privilege by itself?

No. A boundary is a ceiling, not a grant, and not a substitute for a narrow identity policy. The template can require the boundary and still forbid actions in the identity policy. A narrow policy with no boundary, on a role someone else can edit, is drift because the next edit has no ceiling.

What if an SCP already denies the extra admin?

The attach is still drift. The row shows attached and effective separately. Effective can be "denied by SCP" today. The decision is still to remove the attach or to change the template. We do not maintain a forbidden grant because a higher policy currently blocks it.

Can the baseline be generated from current access?

We do not build it that way. Current access is the actual side of the diff. Using it as the template makes every existing grant match by definition. The catalog is the functions the organization can name, changed by pull request. Grants that match no function stay unowned until an owner names the function or the grant is removed.