Skip to content
Least Privilege Principle

The first 90 days should end with one boundary enforced, not a longer list of findings

A first quarter of cloud least privilege work fits the 90-day usage windows provider tools use. Spend month one on logging, data and owners, month two on choosing and simulating one class of access, and month three on audit mode, enforcement and a standing review. Finish with one boundary live and proved.

Open the calculator ยท The field guide

Plan around the 90-day window

The tools that propose removals from usage want about 90 days of history. Google documents a default minimum observation period of 90 days for role recommendations, adjustable to 30 or 60 days for project-level ones. AWS lets Access Analyzer read up to 90 days of CloudTrail events when it generates a policy. If logging is not on today, the first job is turning it on, because every later step depends on the data it produces.

That makes a 90-day program natural. Month one collects and assigns. Month two chooses and tests. Month three enforces and sets the review that keeps the result. A program that tries to enforce in month one is guessing, and one with no enforcement by month three has produced a report. To size the whole clean-up before you plan it, use the Access Debt Calculator and read its assumptions.

Month one: inventory, logs and owners

Confirm that audit logging covers every account, subscription and project, and that it is retained for at least the window you plan to use. Turn on the provider analyzers in report mode: Access Analyzer unused access findings, the Google IAM recommender, and Defender for Cloud permissions management if you run Azure. Produce a first list of unused roles, access keys and dormant identities.

Assign an owner to each account or project. Removal fails when the person who can approve it cannot be found. NIST SP 800-53 control AC-6(7) calls for a review of user privileges, so record who performs that review, how often, and where the result is kept. Keep the list short enough to act on, since a first list of two thousand items teaches nothing. The policy template gives the owner and review wording.

Month two: choose one class and simulate it

Pick one class of access to remove, using the data from month one. Good first candidates are the ones with few legitimate users: unused regions, unused services, access keys unused for the whole window, and roles that trust accounts outside the organization. Choose the one where you can name every principal that currently uses it. Starting small is a choice made on purpose, since the first change builds trust in the process before a larger one needs it.

Write the change as code. Replay recent access under it with the AWS policy simulator, Google's Policy Simulator, or the simulation of any product you use. Review each changed outcome with its owner. A change nobody can explain stays in the branch. Keep a list of the principals that used the access during the window so you can contact each owner before enforcement. The class you pick is a boundary in the sense of what a boundary is.

Month three: audit mode, enforcement and the review

Release the change in audit or log-only mode where the cloud offers one, and watch blocked attempts across a full business cycle. Fix or exempt each hit with a recorded reason. Then enforce, and keep the previous version ready to restore. Write the date and the evidence into the change record, because an auditor will ask for exactly those two things.

Close the quarter by setting a recurring review. Microsoft documents access reviews in Entra ID as a way to review group memberships, application access and role assignments regularly so that only the right people keep access. Add the new boundary to that cycle, and add a pipeline check so the removed access is not granted again. The check keeps a quarter's work from being undone the following week.

What to measure at day 90

Report four numbers: the share of accounts with logging verified, the count of unused roles and keys removed, the number of principals whose outcome changed under the enforced boundary, and the number of exceptions open past their expiry date. A fifth, if you have a reachability model, is the count of paths removed. These are plain counts that any team can produce, and they make a quarter's progress readable by someone outside the team.

If the quarter stalls, why least privilege stalls lists the four usual stopping points. The next quarter repeats month two and three with the next class of access. The routes page lists ways to get help with the work.

Sources

  1. Google Cloud, role recommendations overview, read 16 Nov 2026
  2. AWS, generate policies based on access activity, read 16 Nov 2026
  3. NIST SP 800-53 Rev. 5 catalog (OSCAL JSON), control AC-6, read 16 Nov 2026
  4. Microsoft, access reviews in Microsoft Entra ID, read 16 Nov 2026
  5. AWS, testing IAM policies with the policy simulator, read 16 Nov 2026
  6. Google Cloud, Policy Simulator for allow policies, read 16 Nov 2026
  7. AWS, IAM Access Analyzer findings, read 16 Nov 2026

Last reviewed 16 Nov 2026.