KubernetesSecOps in 90 Days: A Step-by-Step Starter Plan

KubernetesSecOps in 90 Days: A Step-by-Step Starter Plan

You inherited a Kubernetes cluster, or you built one quickly and now people are asking whether it is secure. Good news: there is a sensible order of operations, and you do not need a dedicated security team to follow it. This is a step-by-step walk through the first ninety days of KubernetesSecOps, the daily practice of controlling who changes a cluster, what runs on it, what connects to what, and how you would spot an attack. I will keep each step small enough to finish in a sitting or two.

The outline follows the staged plan in the Ops Atlas page on KubernetesSecOps, which splits the work into three phases it calls see, guard and verify. I have added commands to try and traps to avoid. Use a test cluster first if you can.

Phase one: see what you have (days 1 to 30)

Step 1. Make an inventory

List namespaces, who owns each, and what runs there. Note anything exposed outside the cluster and every component that sits in front of your applications, such as an ingress controller or a service mesh. Record the version and whether the project is still maintained. You will want that list the first time a serious vulnerability headline appears.

Step 2. Replace shared logins with named identities

Connect the cluster to your company’s identity provider using OIDC, so access follows employees joining and leaving. Then check the built-in group called system:masters. Nobody should be in it, because membership bypasses every permission check. If your CI system is using cluster-admin, write that down as a task for phase two.

Step 3. Switch on audit logging

A self-managed API server writes no audit log until you hand it a policy file, and some managed services keep control-plane logs off until enabled. Fix this early; every later decision depends on evidence. Two tips from the Atlas: rules are matched in order and the first hit wins, so put specific rules above general ones, and log Secrets and ConfigMaps at the metadata level so the log never holds their contents.

Step 4. Turn on pod security in warn mode

Kubernetes has a built-in feature, called Pod Security Admission and stable since version 1.25, which you control with namespace labels. Set the baseline level to warn on every workload namespace. Nothing will break, but your deploy output will start showing what would be refused. Collect those messages for a few days and group them by cause: privileged containers, host networking, missing seccomp settings.

Step 5. Run a benchmark and keep the output

A benchmark scanner checks nodes against the CIS Kubernetes Benchmark, which stood at version 2.1.0 in October 2026 according to the Atlas. Save the report with a date. Do not try to fix every line at once; sort failures into “easy”, “risky” and “not applicable to our setup”.

Phase two: guard the doors (days 31 to 60)

Step 6. Enforce the baseline level

Choose a date, announce it, and tell teams the warnings you collected. Aim for most namespaces enforcing baseline; the Atlas uses 90 percent as its threshold for this stage. Two traps to remember. Enforcement looks at pods, so a rule that targets Deployments will not do what you expect, and it never evicts pods that are already running. Remember to schedule a clean-up for them.

Step 7. Keep an exception register

Some workloads legitimately need privileges, such as node agents. For every exception, log a named owner, a justification and a lapse date, then review the list monthly. A related trap: if you exempt a controller’s service account, you implicitly exempt anyone who can create that type of workload, so exempt narrowly.

Step 8. Pilot default-deny networking

Pods accept connections from anywhere until a NetworkPolicy selects them. Choose one low-risk namespace, apply a deny-all rule for ingress and egress, and add allow rules for DNS and for the specific paths your application needs. First confirm that your network plugin enforces policies; kind, for example, has done so since version 0.24.0. Be aware that connections opened before the policy was applied may stay open, depending on the plugin.

Step 9. Encrypt Secrets at rest

Without configuration, Secrets sit in etcd unencrypted. Configure encryption using a key-management-service provider, version 2, since the first version was deprecated in 1.28 and disabled by default from 1.29. The Atlas rates one older cipher mode as weak, so avoid it. Afterwards, rewrite existing Secrets so the old data is encrypted too.

Step 10. Tighten the risky permissions

Search your roles for grants that quietly reach a lot: listing or watching secrets, creating or updating workloads, exec and attach, impersonation, escalate and bind verbs, minting service-account tokens, and approving certificate requests. Replace broad grants with narrow ones. Give the CI system a role scoped to the namespaces it deploys to.

If this feels like the same discipline we use when sizing infrastructure for traffic spikes, it is. Our article on handling large volumes of web traffic on cloud platforms makes a good companion for the capacity side.

Phase three: verify and watch (days 61 to 90)

Step 11. Add a second checkpoint at admission

Pipelines can be bypassed by anyone with kubectl or Helm, so verify at the cluster as well. Start with something stable: images must come by digest rather than by a movable tag. The Atlas warns against blocking every image that a scanner flags, because findings change daily. A policy engine can do this, and the popular ones differ in failure behaviour. As of October 2026, Kyverno fails closed by default while Gatekeeper fails open, so choose the failure policy deliberately for each risk tier.

Step 12. Apply the strict profile to new namespaces

The restricted profile asks for non-root, no privilege escalation, the default seccomp profile and all capabilities dropped. New namespaces are a good place to start since there is nothing to break. Existing ones can follow when the warnings go quiet.

Step 13. Add runtime detection and tune it

Open-source runtime sensors exist for this. Whatever you pick, spend time on tuning. In the Atlas’s worked example, excluding audited on-call sessions took one alert stream from 5.0 percent actionable to 41.4 percent. Send alerts to someone who owns them, and do not let a tool delete pods automatically; that destroys evidence and does not revoke any stolen tokens.

Step 14. Run a drill

Pick a scenario: a leaked token, a compromised pod, a bad admission rule. Walk through it with the team that handles security alerts. Write down what was slow or unclear, then fix those gaps. Two quarters of passed drills is how the Atlas defines its top maturity level.

Keeping it going

After day 90, switch to a rhythm: weekly review of refusals and alerts, monthly exception review, quarterly permission and benchmark review, a yearly tabletop exercise. Before building new features, check the requirements side too; the walkthrough on engineering requirements for web projects helps ensure security needs are written down at the start rather than bolted on.

Common stumbles

  • Switching on every sample policy in enforce mode on day one.
  • Setting every webhook to ignore failures so nothing ever blocks anything.
  • Treating a namespace as a hard security boundary.
  • Trusting a clean permission scan as proof of a clean cluster.

Do the steps in order, date your evidence, and you will be able to answer the next security question with facts instead of hope.