A few years ago a maintenance lead I will call Dana (a composite, not a real person) told her trainees that an audit is just a pre-flight walk with a clipboard and a bit more patience. She said it while holding a binder of paper checklists. Today the equivalent binder lives in cloud dashboards, and the walk-around needs its own clipboard. This post gives you that clipboard: a CloudSecOps self-audit you can run in an afternoon, built from the practices summarised in the Ops Atlas page on CloudSecOps. Mark what holds, ring what does not, and put somebody’s name beside every ring.
Before you begin: set the scene
Pick the scope for this pass. One organisation, one set of accounts, one calendar quarter. Grab whoever administers those accounts and whoever pays the bills, because the second person often knows about forgotten accounts the first has never seen. Bring coffee. The point is honesty, not a perfect score, and Dana always said the audits that went worst were the ones where people tidied the evidence first.
Section A: Do you know what you have?
- A written register names every cloud account, subscription or project, and a real person answers for each.
- Someone has compared that list with what the finance team actually pays for, and the two agree.
- Dormant accounts, such as those tied to former staff or abandoned experiments, are either closed or marked with a reason to stay.
If the second tick is hard, you are in good company. The surprise bill that reveals an unknown account is one of the commonest discoveries in a first audit. Our own note on tracing where monthly spending really goes makes the same point about household money: you cannot protect what you have not noticed.
Section B: Is there a record you can trust?
- Activity logs from all accounts are shipped into a single archive.
- The people who administer the accounts cannot delete or edit that archive.
- You know the retention period for each log source and have extended it where the default is short.
- Data-level logging is on for storage that holds anything sensitive, accepting the extra cost.
Defaults catch people out. Standard event history on one major cloud only goes back ninety days, and detailed object activity is off until you pay to switch it on. Another provider’s sign-in logs last a week on its free tier. In the 2023 hosted-mail intrusion reviewed by a US board, the affected agency lacked visibility into events from before its logging window, which is the audit equivalent of a missing logbook page.
Section C: Are the guardrails real?
- At least one preventive rule blocks public storage and another protects the audit trail from being switched off.
- New rules were tested on a small group before being widened.
- You know which accounts or roles the rules do not reach. On one large platform, organisation-wide policies leave out the management account along with service-linked roles.
- Nobody applied a deny-everything rule at the very top and then left a pile of exemptions behind.
The third tick is where careful teams distinguish themselves. A rule you believe covers everything but secretly skips your most privileged account gives you confidence you have not earned.
Section D: Findings, owners and exceptions
- Every critical finding has an owner, and every owner was assigned within a working day.
- Fixes are confirmed by rescanning, not by a ticket comment.
- The exception register lists reason, owner and expiry for each entry, and none runs beyond ninety days.
- Exceptions are reviewed weekly, and none has been quietly renewed more than once without discussion.
Renewal without review is the sneakiest failure. Dana used to say a waived inspection item that never gets looked at again is simply a deferred accident.
Section E: Keys and sessions
- There is an inventory of encryption keys and what depends on each.
- Your runbook mentions the built-in deletion wait, from about a week to a few months depending on the service, so nobody panics or schedules deletion carelessly.
- The team does not believe that rotating a key re-protects data already encrypted with it. Provider documentation says it does not.
- Your offboarding plan accounts for credentials that stay valid for hours after you disable a user, and closes the route to obtain new sessions first.
Section F: Do the numbers hold up?
- You can state the share of active accounts under current guardrails and which ones are missing.
- You can say how many account and Region pairings have detection switched on.
- You report critical-finding age as a median beside the open findings’ 50th and 90th percentile.
- You treat your tool’s pass rate with suspicion, because some products mark controls passed when findings are suppressed.
- You can name the benchmark version you test against, and you know that a provider’s built-in check may support only an older one.
A quick example of the audit in action
Imagine a small training organisation that runs this clipboard for the first time. Section A turns up a project created for a one-day workshop eighteen months ago, still running and still holding learner spreadsheets. Section B shows logs exist but are kept inside the same account they describe, so a careless administrator could remove them. Section D reveals three exceptions that expired in the spring and were never revisited. Nothing here is unusual. Each item is the kind of thing a ninety-minute conversation with a clipboard can surface, and every item has an obvious first fix.
The emotional effect matters as much as the technical one. People who feared the audit would be a trial usually relax when they see that the findings are mostly ordinary, and the team starts treating the review as routine maintenance rather than judgement.
Scoring yourself
Count the ticks. Mostly ticks in sections A and B and few elsewhere? You are at the early rung, where the aim is simply to see clearly. If sections C and D are solid, you have moved into the baselined and enforced stages of the Atlas ladder. If you have also run a rehearsal of an incident and found the gaps, you are into engineered territory, which very few small teams reach and none need to rush.
What to do with the circles
Pick the three circles that scare you most. Give each an owner, a date no more than thirty days out, and a rescan or a rehearsal as proof. Park the rest on a list with names. Then book the next audit for ninety days from now, because this clipboard only pays off when you use it again.
People newer to the field might enjoy reading our overview of building a technology career through structured courses, which pairs well with an audit like this as a practice project. And if your whole estate is one or two accounts under one team, relax: Microsoft’s guidance supports combining platform and security duties, so Sections A, B and D alone are a strong audit.

