Platforms that coordinate care across borders, handling enquiries, document exchange, scheduling and payments, depend heavily on cloud software. Many of them run on Kubernetes, and the way that layer is secured has shifted a good deal in recent years. This state-of-practice review looks at six shifts in KubernetesSecOps as of late 2026, using the Ops Atlas as the factual backbone and avoiding any clinical or medical-advice content. It is written for technology owners, operations managers and anyone who needs a precise sense of where the practice stands.
A note on terminology first. The Atlas treats the word as an overlapping synonym that is almost unused in the wild; the established name for the activity is Kubernetes security, supported by the project’s own documentation, the CIS Benchmark, the OWASP list of Kubernetes risks and the CKS certification path. Readers wanting the full account can consult the Atlas page on KubernetesSecOps. The six shifts below are my selection and arrangement, not the Atlas’s headline list.
Shift 1: Built-in admission has replaced the old policy object
The legacy PodSecurityPolicy mechanism is gone. Pod Security Admission reached stable status in release 1.25 and gives every cluster three graded profiles out of the box. Since then the API server has gained expression-based validating policies, stable from 1.30, and a mutating counterpart reached 1.36. The Atlas says the project now recommends these expression-based policies wherever they can do the job, which reduces the need to run an external webhook for simple rules. The practical effect for operators is fewer moving parts to keep alive, and fewer places where a webhook outage decides whether requests pass.
One caution survives the change. Published government hardening guidance written in 2022 was framed around the removed policy object, so documents of that vintage need translating before use.
Shift 2: Policy engines are maturing and their failure modes are better understood
External engines remain popular for richer rules. The Atlas observes that one such engine, Kyverno, gained graduated status in the CNCF during 2026, and it records a behavioural difference between engines: Kyverno refuses requests when its webhook cannot be reached, whereas another popular engine lets them through under its default setting, per the October 2026 note. Mature teams now decide the failure behaviour per tier of workload and treat the webhook as critical infrastructure, with its own health checks and redundancy.
Shift 3: Rollouts are staged and exceptions expire
The old habit of switching on a policy bundle and dealing with the fallout has lost credibility. The Atlas describes a promotion path from audit to warning, with a warning period of around two weeks and a published enforcement date, followed by enforcement paired with an exception register in which every entry expires. A governance detail matters here: enforcement does not evict pods that were already running, so a cluster can pass new-workload checks while carrying legacy ones. Teams that measure coverage by namespace, rather than by number of policies, spot that gap sooner.
Shift 4: Secrets handling is moving to external keys and external stores
Defaults continue to surprise newcomers: secrets are stored unencrypted in the cluster datastore until encryption is configured. The Atlas points to the current key-service integration for encryption at rest, noting that its predecessor was deprecated in release 1.28 and switched off by default one release later, and that a basic symmetric cipher option is rated weak. Alongside this, operators that synchronise secrets from an outside vault, and drivers that mount them at start-up, are becoming the normal choice for sensitive credentials. For a platform exchanging personal documents across jurisdictions, keeping credentials and keys outside the cluster shrinks what an intruder could take in one stroke.
Shift 5: Runtime detection is judged by actionability
Runtime detection has also matured, with an open-source detection project graduating from the CNCF in 2024 and in-kernel enforcement tools now able to stop a process mid-misbehaviour. The surrounding discussion has moved from “do we have alerts?” to “how many of them lead to action?”. The Atlas gives an illustration where over twelve hundred monthly alerts yielded only sixty-two actionable ones until expected on-call work was excluded. It also warns against automation that deletes pods on an alert, because it destroys evidence and does not cancel stolen credentials.
Shift 6: Supply-chain incidents changed attention to the components inside the cluster
The retirement of the ingress-nginx controller is the most instructive event of the past two years. A critical flaw announced in March 2025, scoring 9.8, together with four related issues, let an actor on the pod network take over clusters through the admission webhook, and the controller’s default rights reached every secret. Maintainers archived the project in March 2026, so fixes have stopped. Oversight committees warned that, without proactive checking, an affected organisation may only find out once it has been compromised. The shift in attention is toward inventorying third-party components, assigning owners and planning exits before support ends.
What the survey data adds
The Atlas cites a 2024 survey of 600 respondents, self-reported, from a vendor’s research: about two-thirds had postponed a release over security worries, and roughly nine in ten reported at least one incident in the previous twelve months. Runtime issues, build and deployment issues and misconfiguration were the most common categories. Self-reported figures carry the usual caveats about sample and wording, but they match the experience of most platform teams that security work and delivery speed are in constant negotiation.
What stays the same
- Namespaces remain a labelling and policy-scoping tool, not a hard boundary; pods still share kernels, and the stable user-namespace feature in 1.36 narrows but does not remove that concern.
- Access reviews still carry the most value per hour spent, particularly for rights over secrets, workload creation, container sessions and impersonation.
- A clean scanner report still does not prove a clean cluster.
Practical implications for a care-coordination platform
- Inventory every third-party component in the cluster and note its support status.
- Check that secrets are encrypted at rest and that few identities can read them.
- Confirm default-deny network rules for the namespaces that handle documents and payments.
- Ask what happens to admission when your policy webhook is down, and write the answer in the runbook.
- Report coverage percentages to management each quarter.
How to use this review
A trend list is only useful if it feeds a decision. Take each shift and ask one question of your own platform: where do we stand, and who owns the answer? For the admission shift, the question is whether any simple rule still depends on an external webhook that could be replaced by a built-in policy. For the secrets shift, it is whether any credential could be moved out of the cluster. For the supply-chain shift, it is whether every third-party component in the cluster has a named owner and a known support end date.
Record the answers in one table with a date beside each, and revisit the table every quarter. Operations managers in cross-border coordination businesses often juggle several vendors and jurisdictions at once, so a shared page that shows what has been checked, by whom and when, is more valuable than another meeting. It also gives auditors and partner organisations a clear artefact to examine, in place of a verbal assurance.
Related reading
For the engineering context, an enterprise overview of DevOps, Kubernetes and cloud modernization sets out the wider modernisation backdrop. Readers who think about due diligence in general may also like the article on checking accreditation of hospitals abroad, which reflects the same preference for independent evidence that underpins good security reviews.
Looking ahead, expect the trend lines to continue: more logic inside the API server, fewer bespoke webhooks, tighter defaults from managed providers and sharper scrutiny of third-party components. The fundamentals of access, admission, network, secrets and detection will not change, and a team that measures them every quarter will absorb the shifts without drama.