A Practical KubernetesSecOps Audit Checklist for Business Teams

A Practical KubernetesSecOps Audit Checklist for Business Teams

If your company runs software on Kubernetes, someone is making security decisions on your behalf every day, whether they call it that or not. KubernetesSecOps is the shorthand some teams use for that daily work: who can change the cluster, what is allowed to run, what can talk to what, how secrets are stored, and how you would notice an intruder. This post is a plain audit you can run in an afternoon, with no theory beyond what you need to tick a box.

The framing comes from the Ops Atlas, which describes the topic in a long reference entry on KubernetesSecOps. I have turned its main points into a checklist and added a note on what a good answer looks like for each one. If you manage the budget rather than the cluster, hand this to your engineers and ask for evidence, not reassurance.

Before you start: three ground rules

  • Ask for screenshots, command output or exported configuration. “We think it is on” does not count.
  • Write down the date of each check. Reviews that are not dated get repeated or forgotten.
  • Keep a short list of exceptions with owners and expiry dates. You will need it in section 4.

Also set expectations about scope. This audit covers the cluster’s own protections. Building container images, upgrading versions and backing up the data store belong to neighbouring jobs, and your cloud account settings belong to another one still. Mixing them makes the audit endless.

1. Access: who can change the cluster?

Role-based access control, usually shortened to RBAC, is the permission layer. Check these items:

  • People sign in through your company identity provider rather than shared credentials.
  • Nobody is placed in the built-in group that bypasses all permission checks, and your automation systems do not run with full cluster-admin rights.
  • Permissions that quietly reach everything are rare and justified: listing secrets, opening shells in running workloads, impersonating other users, approving certificate requests and minting service-account tokens.
  • Someone reviews all of this at least every quarter.

A good answer is a short, named list. A bad answer is “the CI user has admin because it was easier”. That one sentence appears in a lot of incident write-ups.

2. Admission: what is allowed to start?

Admission control inspects each new workload before it runs. Ask your team:

  • Is the built-in Pod Security feature, or a policy engine, active on every namespace that holds real workloads?
  • Which stage is each namespace at: only observing, warning, or enforcing? A sensible target in the Atlas is enforcing the baseline level in at least nine of ten namespaces, and stricter rules for new ones.
  • Does the checkpoint fail open or fail closed if it goes down? The Atlas notes that two popular engines default differently, so this is a question worth asking by name rather than assuming.
  • Does the team understand that enforcement only affects pods that start afterwards? Workloads already running are not evicted.

Good teams roll rules out gradually: observe first, warn for roughly two weeks with a published enforcement date, then block. If somebody switched on every sample policy in blocking mode on day one, expect chaos and a pile of emergency exceptions.

3. Network: what can talk to what?

By default, pods accept connections from anywhere inside the cluster until a NetworkPolicy selects them. Check:

  • Which namespaces have a default-deny rule? In the Atlas worked example, 31 of 46 namespaces, about two thirds, had one. Know your own number.
  • Your network plugin actually enforces policies. Some setups accept the rules and ignore them.
  • Rules for connections to cloud metadata services and the control plane are explicit rather than left open.

Do not accept “namespaces keep teams apart” as an answer. Namespaces are a labelling and permission boundary. Pods from different namespaces can share a machine and its kernel, so they are not a hard wall.

4. Exceptions: who is excused, and until when?

Every guardrail has waivers. Ask to see the register and check that every entry names who owns it, why it exists and when it lapses. Then look for the two classic traps. First, rules apply to pods, not to the deployments that create them, so a policy written against the wrong object silently does nothing. Second, exempting a controller’s service account effectively exempts anyone who can create that kind of workload.

Review exceptions monthly. An exception list that has never lost a line is a warning sign in itself.

5. Secrets: where do passwords and keys live?

This is the section most business owners are surprised by, so here it is without jargon. Unless encryption at rest is configured, secrets are stored in the cluster database in readable form. The encoding people see on screen is not encryption.

  • Is encryption configured, and through a key-management integration rather than a key sitting in a file next to the data?
  • Does the team use the current version of that integration? The Atlas mentions the first version was deprecated and then disabled by default in 2023 and 2024 releases.
  • Are secrets pulled from an external store for the most sensitive credentials, so the cluster holds as little as possible?

6. Logging and detection: would you know?

Ask three things. Are audit logs switched on for the control plane? On self-managed clusters they produce nothing without a policy file, and on at least one major managed service they are off until enabled. Where do logs go, and for how long? And is there runtime detection that watches for suspicious behaviour inside workloads?

Then ask about noise. The Atlas works through one example where monthly alerts reached 1,240 yet just 62 deserved action; once audited maintenance sessions were filtered out, 58 of the remaining 140 did. If your team cannot tell you the equivalent ratio, the alerts are probably being ignored.

7. Cadence: who looks at this, and when?

A one-off audit decays within months. Put a rhythm on the calendar: a weekly glance at refused deployments and alerts, a monthly exception review, a quarterly check of permissions and a standard benchmark, and a yearly exercise with your security operations people. The Atlas treats 24 months between benchmark runs as the outer limit, not the goal.

Two reality checks

First, patching and attention matter. A widely used ingress component was found in 2025 to have a critical flaw, scored 9.8 out of 10, that let something on the pod network seize a cluster through its admission webhook. The project has since been retired. If a component in your stack is archived, that is an audit finding by itself.

Second, a survey the Atlas cites from 2024 reports that around two thirds of respondents had delayed a deployment because of security concerns, and nearly nine in ten had at least one incident in the previous year. It was self-reported and covered 600 people, so treat it as a mood indicator rather than a measurement. It does show security is a delivery problem, not just a risk problem.

Scoring your results

Give yourself one point for each section where you got evidence, not assurances. Zero to two points means you are on defaults and should start with logging and access. Three to five means you have hygiene and should add enforcement and default-deny rules. Six or seven means you are in a position to start verifying, for instance signature checks and rehearsed drills.

Wondering where this sits in your wider technology plan? Our piece on choosing and steering business technology services covers the buying side, and the article on growth strategy for product-led companies is useful for deciding how much delivery speed you can trade for safety. Then book the next review before this one ends.