If your hospital’s reporting tools, scheduling portal or document archive now live in a cloud account, somebody has to keep the settings honest. That ongoing upkeep is what people mean by CloudSecOps, and the friendliest way in is a self-audit you can run with a coffee and a shared screen. Below is a checklist for a small administrative team. It is about accounts, permissions, logs and keys only, with no clinical content and no description of any specific hospital.
Work through it once, mark each line yes, no or unsure, and treat the “unsure” pile as your real to-do list.
How to run the session
Book ninety minutes and invite three kinds of people: someone who administers the cloud accounts, someone who owns a system running in them, and someone from operations or procurement who signs the contracts. Nobody should read from a script. Ask the questions aloud, let the room argue, and write down what is said rather than what sounds good. The arguments are the useful part, because they show where responsibilities blur.
Part one: know what you are guarding
- Can we point to an up-to-date roll call of all cloud accounts and the regions each one touches? Include test and old project accounts, because forgotten ones are the usual hiding places.
- Is each account tied to a named owner? A shared mailbox does not count as a person.
- Do we know which duties belong to the provider and which to us? Hardware and data-centre buildings are theirs. Who may sign in, what is publicly reachable and what is stored remain ours.
The Ops Atlas page on what CloudSecOps covers and where it stops presents it as a freshly coined phrase that no standards body has formalised, so expect when two suppliers describe it differently. Request a one-page scope statement from each, and compare them side by side.
Part two: logs you will want later
- Are audit logs switched on in every account? Not only the busiest ones.
- Do we know how long each log source keeps its records by default? Several services keep roughly ninety days, and some record types cost extra to enable.
- Are copies sent to a separate archive that daily administrators cannot edit or delete?
- Has anybody ever tried retrieving an old entry? A log that nobody can read is decoration.
A tip from experience: ask for a test retrieval during a calm week, not during an incident. It takes twenty minutes and reveals whether the pipe is connected.
Part three: public exposure and drift
- Is anything publicly readable on purpose? If so, is the reason written down?
- Is the organisation-wide public-access block in place, or only bucket-by-bucket settings? The Atlas calls out the second approach as brittle because a single edit can quietly undo it.
- When someone fixes a setting by hand in the console, does the template that built it get updated too? Otherwise the next automated run reverses the fix and the problem returns.
- Are any role names exempted from a rule without protecting those roles? An exemption that anyone can impersonate is an open door.
These items look small, but they are the ones that turn up again and again when a review of a past cloud incident gets published. The Atlas also notes that, in an expert survey published in 2024, misconfiguration and weak change control ranked first among cloud threats, with identity problems second. It is an opinion poll of specialists, not a count of real incidents, so quote it gently.
Do not worry if the room goes quiet on several of these lines. A hush usually signals a duty that everybody assumes somebody else handles. Assign each quiet line to a named person before leaving the meeting, and put the date of the next review in the notes.
Part four: people and credentials
- Is a second sign-in step required for every human account?
- When did long-lived access keys last get replaced? Stale keys are the quiet danger.
- If we had to cut off a compromised session today, do we know how? Short-lived credentials can stay usable well past a working shift in some setups, so ending one session may fall short.
- Do we have an inventory of encryption keys and their owners?
One myth worth ruling out early: regularly rotating an encryption key does not, on its own, change who can read the data that key protects. Rotation is housekeeping, not a shield. Good practice pairs it with tight permissions and with clear waiting periods before a key can be destroyed.
Part five: findings and exceptions
- Is a named person attached to each serious finding by the next business day?
- Is closure confirmed by scanning again?
- Do exceptions carry an expiry date, and does the team actually review them? The Atlas proposes three months as a sensible upper limit.
- Do we track how many exceptions have already expired? Divide expired by open to get a quick honesty score.
Ready to score yourself? Count the yes answers. Under half means the first thirty days should be about inventory, logging and owners. Somewhere between half and three quarters points to focusing on exceptions and staged rollouts. Above that, schedule a rehearsal where the team talks through a pretend control-plane incident and notes where the instructions fall short.
If you also want a sense of where you stand over time, the Atlas sketches a five-rung ladder that starts at improvised work and finishes at routinely tested detection. Nobody needs to chase the top rung. Simply reaching the second one, where each critical item has someone responsible, is a real gain for a small team.
It is candid about when this is overkill too. A hospital with a couple of accounts and one technical team may do better by merging platform and security responsibilities than by creating a separate function. Decide candidly where yours sits.
Part six: automation with brakes
- Which fixes are automatic today? Blocking something risky from being created is a safe candidate.
- Could each automatic repair be undone, and is it narrow enough to be predictable?
- Is there a rule that scripts never delete stored data and never rotate credentials that are in use?
Tools that help here include Prowler, an open-source scanner with an Apache licence, and Cloud Custodian, an incubating project in the Cloud Native Computing Foundation. Neither is required. What matters is that a person chooses the policy and knows what the tool will touch.
A weekly and monthly rhythm to adopt
Checklists fade unless they are attached to a calendar. A sensible pattern is daily attention to critical items, a weekly look at exceptions, a monthly review of the rule set and benchmark versions, and a quarterly rehearsal. Put the dates in the same shared calendar your team already uses for maintenance windows.
For more background on tidy cloud foundations, our note on keeping cloud infrastructure manageable pairs well with this one, as does the article about modern delivery habits for cloud operations explains the engineering context around these checks.
Wrapping up the audit
Print the list, answer it honestly and share the unsure items around, each tagged with an assignee and a due date. If a supplier offers a shiny dashboard instead of answers to these questions, thank them and keep asking. A good dashboard is a summary of work that was already done properly, not a replacement for it.
