This is a working checklist, not an essay. It is written for anyone at an aviation training operation who has been handed responsibility for the systems behind the scheduling, records and simulator booking tools, and who has been told those systems “run on Kubernetes”. The audit below follows KubernetesSecOps, the routine work of keeping a cluster protected, and it is arranged so you can go through it in an afternoon with a notepad. Every item carries its own pass condition. If you cannot answer the question, that is itself the finding.

Before you start: KubernetesSecOps scope and sources

Decide what you are auditing. The Ops Atlas defines KubernetesSecOps through five concerns: control over cluster changes, control over admitted workloads, control over connections between pods, handling of credentials, and spotting an attack. It does not cover building container images, upgrading the cluster, cloud account settings or incident triage, which belong to neighbouring disciplines. The Atlas page on auditing and running KubernetesSecOps is the source for the checks here, so keep it open next to this list.

Think of a training school’s compliance review: you do not check the runway surface during a records audit. Same discipline, same boundaries.

Section A: Access (who can change the cluster)

  1. Login method. Is sign-in handled by one central identity provider, or by shared static credentials passed around? Pass: central sign-in, no shared admin accounts.
  2. Broad roles. List every identity holding cluster-wide admin rights. Pass: a short list, each name justified.
  3. The bypass group. Confirm nobody sits in the special system group that bypasses rights checks altogether. Pass: empty, except for emergency break-glass that is logged elsewhere.
  4. Risky verbs. Search roles for the dangerous permissions the Atlas names: listing or watching Secrets, creating workloads, opening shells in running pods, impersonating, approving certificate requests, minting service-account tokens, and editing admission webhook settings. Pass: each is held by a named, reviewed role.
  5. Automation accounts. Does the delivery pipeline run with cluster-wide admin? Pass: no; it holds only what its deployments need.

Record dates of the last review. The Atlas suggests a permission review every quarter, with the outer limit between full reviews put at two years. Think of this as the interval for a mandatory inspection: the outer limit is not the target.

Section B: Admission (what may run)

  1. Baseline pod rules. Are the basic pod safety rules in warn mode or better on every workload namespace? Pass: yes, with a date to enforce.
  2. Stricter profile for new areas. Do new namespaces start on the restricted profile: non-root, no privilege escalation, default syscall filter, all capabilities dropped? Pass: yes.
  3. Failure behaviour. For each policy engine, what happens when it cannot be reached? According to the Atlas, as of late 2026 one widely used engine rejects requests by default when unreachable, while another lets them pass. Pass: you know the answer for yours, and it is set per workload tier.
  4. Exception register. Is there a written list of exemptions, each with a reason and an expiry? Pass: yes, reviewed monthly.
  5. Exemption reach. Does any exemption name a controller’s service account? If so, remember it also exempts anyone able to create that workload kind. Pass: reviewed and accepted in writing.

A reminder for the inspection notes: enforcement applies to new requests only. Pods already running when a rule was switched on are untouched, so schedule a one-off review of what is already live. The same applies to equipment: aircraft already delivered still need a walk-through whenever new gear arrives, as the post about the glass-cockpit twin added to the training fleet implies.

Section C: Network (what may talk to what)

  1. Default deny. Which namespaces have a deny-all network policy? Count them. Pass: every namespace that holds student, payment or scheduling data.
  2. Test the policy. Does your network plugin actually enforce policies? Test from a throwaway pod. The Atlas notes that even local test environments only began enforcing them in recent versions.
  3. Old connections. After enabling deny-all, restart workloads so previously opened connections do not linger. Pass: done and noted.

One note on how to test the network step honestly. Create a short-lived pod in a namespace you believe is locked down, try to reach a service in another namespace, and record whether the attempt succeeded. A pass on paper means little if the plugin underneath does not enforce policies. Do it twice: once before and once after the deny-all rule, and keep both results in the audit file.

Section D: Secrets and storage

  1. Encryption at rest. Is Secret encryption configured on the control plane, using the current key management approach? The Atlas says the older version is deprecated and disabled by default in later releases, and rates one cipher option as weak. Pass: configured, verified by reading a stored Secret directly.
  2. Base64 is not protection. Does anyone in your team treat base64 values as hidden? Pass: nobody does.
  3. External secret stores. Where possible, are credentials pulled from a dedicated store instead of living in manifests? Pass: yes for payment and records systems.

When checking encryption, do not stop at the configuration file. Fetch a stored Secret straight from the datastore using an administrator’s tools and confirm that what comes back is unreadable. This single extra step separates a configuration that exists from one that works, and it takes about ten minutes.

Section E: Logs and detection

  1. Audit policy. Is one loaded? Without it, a self-run control plane keeps no trail. Pass: loaded, with Secrets and ConfigMaps at metadata level.
  2. Managed service defaults. If you use a hosted cluster, confirm control-plane logs are enabled. Per the Atlas’s late-2026 snapshot, one large provider leaves control-plane logging switched off until requested; another always retains admin activity but ships with data-access logging off.
  3. Runtime alerts. Is something watching running workloads, and has anyone measured how many alerts are actually actionable? Pass: yes, with a tuned rate.
  4. Benchmarks. Has a scan against the CIS Kubernetes Benchmark been run, using a tool such as kube-bench? Pass: yes, results stored. Treat a clean scan as a prompt for questions, not proof, since scanners can miss permission routes through operators.

Scoring your afternoon

Tally the passes. There is no magic number, but the maturity ladder in the Atlas offers a frame. If most of Sections A and E fail, you sit on the bottom rung of the Atlas ladder, and its first 30-day stage applies: inventory, warn mode everywhere, audit logging and a first benchmark. If A through C pass and you hold an exception register, you are at guardrails and should look at signature verification and drills. Compare your simulator booking and records systems to a training aircraft: the minimum equipment list tells you what may be missing and still allow flight, and the rest are defects that must be tracked.

Keep the paperwork habit going. Pair the audit with a visit to the guide on how a flight simulator was brought into the IFR training programme to see how new training capability gets documented before it is used.

What auditors usually find

Patterns repeat from one organisation to the next. The most frequent finding is a pipeline account with far more power than its job needs, usually granted in a hurry during setup and never trimmed. The second is a policy engine installed in permissive mode and never moved forward, so it logs problems nobody reads. The third is missing paperwork: no exception register, no record of who approved what, no date for the next review. None of these requires special tooling to fix. They require someone to own the list and a calendar entry. If you are the person holding the notepad, write names beside every failed item, because an unassigned finding tends to stay open indefinitely.

Three follow-ups worth putting in the diary

  • Next week: a quick scan of rejections and alerts.
  • Next month: read the exception register and remove anything expired.
  • Next quarter: repeat sections A and D, and run a benchmark again.

Print the checklist, date it, sign it and file it with the other compliance records. An audit that nobody can find six months later is barely an audit.

© 2024 ADVANCED INTERNATIONAL AVIATION ACADEMY. ALL RIGHTS RESERVED.

STAY CONNECTED WITH US: