Every cluster has a checklist. Most of them are correct, mostly complete, and almost entirely unused. The list is signed off in a meeting, the Confluence page goes stale, the cluster drifts, and the auditor finds the gap eighteen months later. The list is not the protection. The work between the items on the list is the protection.
Kubernetes security in 2026 has settled into a familiar shape. Day one items that have to be right from the first cluster, a second wave that takes months to roll out, and a third wave that has to be in place before a SOC 2 lands. Day one is non-negotiable, day one is also where the budget is most often cut because day one is the work the platform org has to fund before there is anything running. Here is what each layer actually requires.
Day one, in order of pain when missing
Identity, network policy, secrets, image provenance, admission control. Five items, each with a toolchain that ships in a day. RBAC mapped to real roles, not to service accounts called “default”, and ClusterRoles trimmed so a compromised pod cannot list every secret in the cluster. Network policy enforced as default deny, with explicit allow rules for the namespaces that need to talk. Secrets stored in HashiCorp Vault or AWS Secrets Manager, never in environment variables, rotated on a schedule the platform org can defend in a tabletop.
Image provenance is the under-appreciated one. Build the image from a reproducible build, sign it with a verifiable key, deploy through an admission controller that rejects unsigned images. Kyverno, OPA, Connaisseur, any of them. A cluster that accepts any image from any registry is a cluster that a poisoned container can land in. The supply chain attack on the cluster costs less than the audit response later.
The next five: the work that takes longer
Runtime threat detection with Falco, Tetragon, or Tracee. Admission control policies written down and enforced, not documented. Resource limits and quotas on every namespace, so one misbehaving pod cannot take the production offline. Audit logging shipped to a place the SOC actually queries, which is not the cluster itself. Backup and recovery that is tested quarterly, not annually, and tested against a ransomware scenario, not against a happy path.
Each of these takes a quarter, not a sprint. Each of these also fails quietly when it is not done, which is the dangerous version. A backup that has never been restored is not a backup, it is a bill for storage. An admission controller in audit mode is a suggestion, not a control. The platform org has to commit to the rollout calendar before the work is worth starting.
What to do this quarter
Pick the five day one items, get them right on the next new cluster, and migrate the existing clusters one namespace at a time. Day one is a sprint, not a project. The platform org that does day one in two weeks has the budget and the political capital to fund the next five. The platform org that treats day one as a roadmap item is the platform org that ships day one in two quarters, by which time the auditor has already asked.
Then add one item from the next five per month. Resource limits first, because limits are the cheapest control to add and the easiest to defend. Backup and recovery second, because backup is the difference between a recoverable incident and an existential one. The platform org that does this honestly for six months is in a different position than the platform org that did it last year, and the auditor can tell.
The bottom line
Day one in a sprint, next five over two quarters, audit logging shipped somewhere the SOC can query it. The cluster is not protected by the checklist, it is protected by the work the checklist forces the platform org to fund.
Sources & Further Reading
All claims in this article are sourced from primary documentation, vendor advisories, and reputable security researchers.
Spotted an error? Email the editor. Corrections are issued with a visible correction note.
Editorial standards. Every article on humanrequired.org is reviewed by a human editor before publication. AI may assist with drafting or research; final editorial control is human. Read the full standards.
