Kubernetes v1.36 Arrives with Security Hardening, SELinux Acceleration, and Expanded AI Hardware Support
Kubernetes v1.36, releasing April 22, brings accelerated SELinux volume labelling to GA, expands Dynamic Resource Allocation for AI hardware, and permanently removes the insecure gitRepo volume driver.
Overview
Kubernetes v1.36 ships on April 22, 2026, the second release of the calendar year following v1.35 in December 2025. As previewed by the Kubernetes release team, the release is packed with enhancements and places particular emphasis on security hardening and expanded support for AI and machine learning infrastructure. It is notable for what it removes as much as what it adds: the gitRepo volume driver is permanently disabled, ending a deprecation that began in v1.11, and the externalIPs Service field receives its first formal deprecation warning.
The v1.35 release, covered by InfoQ at its December 2025 launch, introduced in-place pod resizing and AI-optimized scheduling improvements; v1.36 continues that trajectory by advancing the Dynamic Resource Allocation framework and tightening the cluster’s security primitives.
What Graduates to Stable
Among the highlights graduating to General Availability, accelerated SELinux volume labelling is the most operationally impactful for Linux-enforcing environments. The improvement, highlighted in the release sneak peek, replaces recursive inode-by-inode relabeling — which on large volumes could consume minutes and materially delay pod startup — with a mount -o context=XYZ approach that applies SELinux context at mount time. Platform teams enable the feature via spec.SELinuxMount on the pod. The release team notes a caveat: pod authors who share volumes across multiple pods with differing SELinux policies must review their configurations carefully, since the mount-time approach surfaces labelling conflicts that the slower recursive path previously masked.
External signing of ServiceAccount tokens also reaches stable in v1.36. This feature delegates the cryptographic signing of service account tokens to external systems — cloud provider key management services or hardware security modules — rather than relying on the static signing key held by the API server. For organizations subject to strict key custody and compliance requirements, this closes a gap that previously required architectural workarounds such as manually rotating API server signing keys and coordinating the associated credential refresh across all service accounts.
Dynamic Resource Allocation Expands AI Hardware Support
Kubernetes v1.36 advances the Dynamic Resource Allocation framework, which provides Kubernetes-native primitives for allocating specialized hardware such as GPUs, FPGAs, and AI accelerators. As the sneak peek post describes, this release introduces device taints and tolerations for DRA — mirroring the familiar node taint model — allowing operators to mark individual accelerators as degraded or under maintenance without cordoning an entire node. Pods that cannot tolerate a device taint will not be scheduled onto that hardware, while other pods continue running on the remaining healthy devices on the same machine.
This granular device-level control is a meaningful operational improvement over the previous model, where a failing GPU typically required taking the entire node out of the scheduling pool. As AI inference and training workloads increasingly run on densely GPU-populated machines, the ability to isolate individual device failures without disrupting co-located workloads reduces operational overhead and improves cluster utilization.
DRA was introduced in v1.33, and each subsequent release has expanded its device model. V1.36’s additions signal the framework is maturing beyond proof-of-concept and toward production-viable accelerator management.
Removals and Deprecations
The gitRepo volume driver is permanently disabled in v1.36. The plugin, which allowed the kubelet to clone a git repository directly onto a node at pod startup time, was deprecated in v1.11 due to a critical security issue: it created a code execution path on cluster nodes through repository contents. As the Kubernetes release team noted, “the kubelet’s job is not to reach out to the internet and clone repositories.” Teams still relying on gitRepo volumes must migrate to init containers or an external git-sync sidecar pattern before upgrading. Any pod spec referencing the plugin will fail to schedule on a v1.36 cluster.
The Service.spec.externalIPs field receives a formal deprecation warning beginning in v1.36, tied to CVE-2020-8554. That vulnerability identified that the field can be abused to mount man-in-the-middle attacks against cluster traffic by routing through attacker-controlled IP addresses. Full removal is planned for v1.43. Administrators relying on externalIPs for direct service exposure — a pattern common in bare-metal deployments — should begin evaluating migrations to LoadBalancer services, NodePort, or the Gateway API.
Operational Context
V1.36 is a pragmatic, low-risk upgrade for most clusters. The two breaking changes — gitRepo removal and IPVS kube-proxy mode removal — have both carried deprecation warnings across multiple prior releases. Platform teams that track the Kubernetes deprecation guide will not be surprised by either.
As InfoQ observed in covering the Kubernetes 1.35 release, the upstream project follows a disciplined deprecation cadence that gives operators meaningful lead time before removals take effect. The pattern continued here: the gitRepo plugin was first deprecated in v1.11 (2018), has been discussed in every subsequent release note, and is only now being hard-disabled.
The DRA device taints feature marks a broader shift in Kubernetes’s posture toward AI infrastructure. Previous releases required operators to rely on vendor-specific device plugins, node feature discovery labels, and external admission webhooks to manage GPU heterogeneity across a cluster. The emerging DRA model replaces this fragmented approach with a unified, extensible device API that the scheduler understands natively. V1.36 represents an inflection point in that evolution: the primitives are now stable enough for production pilots, even if the ecosystem of hardware vendors supporting DRA drivers is still maturing.
What We Don’t Know
The official release blog post on kubernetes.io — which will include the complete list of enhancements, their associated Kubernetes Enhancement Proposal numbers, and the full upgrade guide — had not been published at the time of writing. The sneak peek post noted that information was subject to change before the release cut, and the final scope of several alpha features was still under discussion.
The timeline for Kubernetes v1.37 has not been formally announced, though the project’s three-releases-per-year cadence suggests it would arrive in late summer or early autumn 2026.