If your safety management system lives in software, then the question of who can edit a hazard report, close a corrective action or change a risk rating is itself a safety question. A tidy audit of those permissions, the kind that technology teams file under AccessOps, takes an afternoon and pays back every time an auditor or investigator asks, “who changed this?” What follows is a relaxed, tick-box walk-through for a quiet afternoon, with a printout and a notepad to hand. You need no special software.
Before you start: set the scope
Pick the systems that hold safety-critical records: your reporting portal, the corrective-action tracker, the document library for procedures, and any dashboard that feeds management reviews. For each one, you are looking at the special powers, the ones that go beyond ordinary reporting and reading. The Ops Atlas entry on AccessOps and how elevated rights are requested and retired draws the same boundary: it concerns how extra authority is obtained and withdrawn, not who is on the staff list or how normal roles were designed.
Give yourself a simple sheet with columns for system, power, holder, reason, expiry and approver. You will fill it in as you go.
Part A: Find out who holds what
- [ ] Export or screenshot the list of administrators for each safety system.
- [ ] Mark any shared or generic logins (for example “sms-admin”). Each one is a finding.
- [ ] Mark accounts of people who have left, moved role or gone on long leave.
- [ ] Count the total of always-on powerful accounts and jot that figure across the head of the sheet as your starting baseline.
A large number on the first pass is normal. Most organisations turn up more than they expected, and finding the surplus first is exactly what the exercise is for.
Part B: Check that every power has a story
- [ ] For each holder, can you state in one sentence why they have the power?
- [ ] Is there a recorded expiry? For ordinary elevated requests the Atlas suggests a default of roughly three to six months, with shorter spans for more sensitive tiers.
- [ ] Is there a named approver, or a clear rule that approved it?
- [ ] Is the reason specific (“close audit finding 14”) rather than generic (“needed for work”)?
Blanket approvals are a trap. The Atlas observes that when managers approve everything, justifications become boilerplate and the approval stops meaning anything. If you read ten reasons in a row and they are all identical, the process is decoration.
Part C: Test whether expiry really ends access
Many teams jump over this step, though it exposes the most. Ask the system administrator to remove a test person’s elevated role, then check whether that person’s already-open session can still edit a record. The Atlas explains that a grant and a session are different things: in its experiments a live connection outlasted the credential that allowed it, cloud role sessions run until their own expiry, and permission changes can need minutes to take effect.
- [ ] After withdrawing a right, did the person’s open session stop?
- [ ] If not, is there a way to force sign-out of all sessions?
- [ ] Is anyone able to edit permissions themselves and quietly keep a copy of their own access?
Part D: Examine the emergency route
- [ ] Is there an emergency account for when the normal approver is unreachable?
- [ ] Are there at least two, so one lost phone does not lock everyone out?
- [ ] Does using it raise an alert to a named person?
- [ ] Has someone actually tried it in the past quarter? The Atlas relays a big cloud vendor’s recommendation to prove those accounts still work roughly every quarter.
- [ ] Was the last use reviewed by the next business day?
One trap the Atlas highlights: if the list of approvers is empty or points at people who have left, nobody can approve anything and the whole tenant can lock up. Check that your approver lists contain living, reachable humans.
Part E: Look at tiers and time limits
- [ ] Have you sorted your systems into tiers by consequence?
- [ ] Do the most sensitive systems have the shortest grants, measured in hours, not weeks?
- [ ] Do lower tiers still get an annual review at minimum, as common control benchmarks recommend?
- [ ] Is anyone treating the yearly review as the main control? Reviews alone sit at the earliest maturity stage in a government zero-trust model the Atlas cites.
Part F: Count what really gets used
- [ ] For the last month, how many elevated grants were approved?
- [ ] How many were never used?
- [ ] Roughly how much of the granted time involved real work? In the Atlas worked example, a quarter did, which is a sign defaults were generous.
- [ ] Do dormant accounts, say untouched for six weeks, get flagged for removal?
Part G: Watch the automation
- [ ] If any approval is automatic, who checks the rule? Per the Atlas, a bot that approves requests is as sensitive as the systems it protects, so it belongs in the highest tier itself.
- [ ] Does any infrastructure-as-code or sync job overwrite manually granted bindings without warning?
- [ ] Has anyone proposed letting an AI model approve high-tier requests? The Atlas notes there is no documented safe design for it; summarising for a human is the most it supports.
Part H: Walk through one real change
Finish the audit by picking a single recent change, such as a corrective action closed last month, and tracing it end to end. Who made the edit? Under which account? Did that person hold elevated rights at the time, and if so, who granted them and for how long? Could you show an auditor the answers in under ten minutes?
- [ ] The account that made the change is a named individual, not a shared login.
- [ ] The elevated right, if any, was active only for the task window.
- [ ] The reason in the grant matches what was actually changed.
- [ ] You can show the trail without asking the vendor for help.
Safety culture depends on people trusting the record. If tracing one change takes an hour of guesswork, tell the team honestly, then fix the weakest link: usually a missing reason, a shared account, or logs that nobody retains long enough.
Scoring your afternoon
Count the unticked boxes. Under five, you are in good shape and should schedule the same check six months from now. Between five and fifteen, pick the three most serious gaps, often shared logins, missing expiries and untested emergency routes, and fix those first. Above fifteen, do not panic; instead, adopt the thirty-day plan the Atlas suggests, which begins with inventorying the most sensitive systems and rehearsing the emergency path.
Keep it going
An audit is a snapshot. The habit that matters is a short recurring review plus a rule that every new elevated right arrives with a reason and an end date. Put a reminder in the calendar now. And remember what the Atlas is candid about: this label is editorial rather than an industry standard, but the practices beneath it are well established. Use them to make your safety records as trustworthy as the aircraft they describe.
