How to Run an IAMOps Permission Review in Six Evidence-Led Steps

How to Run an IAMOps Permission Review in Six Evidence-Led Steps

Avionics shops do not trust a unit because it was fine last year; they bench-test it against a written procedure and record what they find. Cloud permissions deserve the same treatment. A review that starts from evidence, not from hunches, is the practical core of IAMOps, a loose label the Ops Atlas applies to authorisation written, tested and tuned like software, with rules for every human, workload and automated agent. Here is a six-step routine a flight-training operation can run on its own scheduling, records and simulator-telemetry systems.

Step 1: Draw the inventory before you touch anything

List every role, user, service account and automation job that can reach systems holding trainee records, aircraft hours or payment data. Next to each, write an owner. A permission without an owner is like a tool without a tag in the hangar: nobody can say whether it should still be there.

Money habits make a surprisingly good warm-up for this stage. The mindset in a walkthrough on plugging small financial leaks transfers directly, because idle permissions are leaks too: individually small, collectively expensive, and easy to ignore until someone reads the ledger.

Step 2: Ask who can reach you from outside

Before pruning internal rights, check for access that crosses your boundary, such as a storage bucket shared with an outside account or a role that a third party can assume. The Atlas notes that external-access analysis on AWS is free, which makes it an easy opening move. Anything surprising here outranks every other finding on the list, so fix it first.

Step 3: Measure utilisation, not just existence

Now set each role’s granted actions beside the actions it has really performed. The Atlas gives a worked illustration: one role with 180 permitted actions that exercised only 27 scores 15 percent. A set of 120 roles had a median near 22 percent, and a quarter of tidy-up lifted it to 41.

The unused-access feature on that same cloud costs twenty cents monthly for every role or user it scans. For 850 roles that comes to about 170 dollars monthly, according to the Atlas’s own sums. Compared with the cost of a data exposure, it is a modest line item, though your own bill depends on the current price list, so confirm it before budgeting.

Keep the blind spots in mind. A quarterly export job will appear dormant in a month-long window. Role-passing permissions and direct data-layer calls are also easy to miss. Mark anything seasonal before it lands on a removal list, the way a maintenance planner flags calendar-based inspections apart from hours-based ones.

Step 4: Cut in small, reversible slices

Remove permissions in batches that you can roll back, and tell the owner first. Where possible, replace stored keys with short-lived federated credentials so that a leaked secret expires quickly. The Atlas points to a worrying pattern in public survey data: in a late-2025 cloud security report, active keys older than a year turned up on 59 percent of AWS IAM users and 55 percent of GCP service accounts. Aging credentials are the norm, which is why retiring them is rarely wasted effort.

If your organisation plans heavier infrastructure work, the engineering fundamentals in a guide to building infrastructure skills will help your team write these changes as code rather than clicking through consoles.

Step 5: Test guardrails in a sandbox, and watch for silent failures

When you add a rule that blocks something, rehearse it somewhere harmless first. Two subtle traps are worth knowing. First, conditions that test for a missing value behave differently depending on whether they are positive or negated: the Atlas warns that a negated operator can match when the key is absent, a result most reviewers would not predict. Second, error handling varies by engine. In one policy language described there, a rule that errors out is silently passed over, leaving a broken blocking rule to fail open. Validating policies against a schema before release closes that gap.

Think of it as a simulator check ride. You want to learn how the rule behaves in an emergency before real trainees depend on it.

Step 6: Record the numbers and schedule the next pass

Close the loop with figures that can be compared next time: utilisation percentage, the share of statements using wildcards, credentials older than ninety days, denials observed after changes, and approval delays. The Atlas’s example shows wildcard exposure of 168 allow statements out of 1,400, or 12 percent. It also reports that grant time rose from 36 to 48 hours after tightening and was accepted by the team as a fair trade.

Put a date on the next review. The Atlas suggests monthly cadence once a team reaches its fourth maturity level, labelled Measured. Before then, quarterly is a sensible rhythm for a small organisation.

Common questions from analytical teams

Do we need a special policy language? Not at the start. Engines such as OPA or Cedar help once your rules live in a repository and pass through automated checks, but a spreadsheet and disciplined review already beat guesswork.

Who should own the exercise? Whoever is accountable for the system, with security acting as a reviewer. The reverse rarely works, because the people running the training platform know which duties are seasonal.

What does the Atlas count as success? Its top maturity level describes removals that cause no incidents. That is a long-run target, not a month-one expectation.

Where to read the primary source

The routine above borrows its shape from the Ops Atlas IAMOps entry, which adds a thirty, sixty and ninety day plan, a note on standards such as NIST SP 800-53 control AC-6, and a candid warning that the label is informal. Read it with your own system list open beside it; the value comes from mapping its ideas to your environment, not from copying its tables.

The final test of a permission review is the same as an aircraft release: could you explain, with records, why each item is where it is? If the answer is yes for every role and key, you have done the work.