Google Kubernetes Engine: security advisories, deprecations and releases

GKE Private Service Connect API timeouts from three-layer networking bugs

GKE clusters using Private Service Connect in europe-west experienced intermittent API server timeouts caused by packet reordering, Envoy half-close behavior, and a Linux kernel TCP bug in FIN_WAIT_2 state handling.

Updated Container-Optimized OS images fix vulnerabilities in GKE

New GKE versions include updated COS images with cumulative security fixes.

GKE Container-Optimized OS security updates

Updated COS images in multiple GKE versions fix cumulative security issues. Check release notes per version.

GCP-2026-071: emSign CA compromise in GKE

Multiple emSign subordinate CA keys were stolen, exposing external TLS connections from GKE nodes and control plane to MitM. Internal cluster mTLS is safe. Patched GKE versions now distrust the bad CAs.

CVE-2026-111943

GKE Cloud Storage FUSE CSI driver issue resolved

Resolved Cloud Storage FUSE CSI driver bug for ubuntu_containerd and older cos_containerd node pools in GKE.

GKE Sandbox Pods stuck in Terminating with Cloud Storage FUSE CSI driver

Pods using Cloud Storage FUSE CSI driver in GKE Sandbox may remain stuck in Terminating state indefinitely.

GKE security updates via new COS images

Updated Container-Optimized OS images bring cumulative security fixes for GKE versions 1.31-1.38; see COS release notes for specific CVEs.

GKE 1.36.3+ FUSE CSI DaemonSet crash on older kernels

Control planes 1.36.3-gke.1244000 and later break the gcsfusecsi-node DaemonSet on Ubuntu and older COS node pools because /proc/sys/fs/fuse is absent.

GKE nodes: updated Container-Optimized OS images fix security issues

Multiple GKE versions ship refreshed COS images containing the latest security patches from upstream COS releases.

Identity Service for GKE deprecated in GKE 1.36 and earlier

Identity Service for GKE will be removed from GKE 1.36 and earlier starting July 1, 2026, and is unavailable for orgs created after July 1, 2025. GKE 1.37+ will not support it.

GKE node upgrades now reuse existing Compute Engine VM

GKE Standard CPU node pools can now upgrade nodes in place by updating the existing VM instead of deleting and recreating it, preserving compute capacity during upgrades.

GKE 1.36 changes kube-dns default to CoreDNS, alters pod IP DNS behavior

GKE 1.36 switches kube-dns to CoreDNS. Pod IP queries with non-existent namespaces now return NXDOMAIN instead of skipping namespace checks.

VPA and HPA can jointly right-size CPU requests

Vertical Pod Autoscaler can now optimize CPU requests for workloads that also use Horizontal Pod Autoscaler on GKE 1.36.3+.

GKE CPU startup boost in preview

GKE now offers CPU startup boost via Vertical Pod Autoscaler. It temporarily raises CPU for faster pod initialization then scales back automatically without restarts.

GKE Agent Sandbox generally available

Production-ready sandboxing technology for secure multi-tenant agent execution on GKE.

GKE HPA supports custom PromQL metrics

Scale workloads using custom Prometheus metrics in addition to standard metrics.

GKE now supports native scale-to-zero with HPA

HPA with Autoscaling Metric and KEP-2021 enables built-in scale-to-zero without extra configuration.

GKE Agent Sandbox for RL generally available

Optimized sandboxing and orchestration SDK for large-scale reinforcement learning loops on GKE.

GKE Pod snapshots reduce AI inference cold starts by up to 89%

Save and restore CPU/GPU memory state of pods to cut inference startup latency.

GKE Dataplane V2 supports 15 000 nodes with NetworkPolicy GA

Standard clusters can now scale to 15 k nodes while keeping full NetworkPolicy enforcement for large AI estates.

gVisor sandboxes for Ray clusters on GKE in preview

Experimental library brings gVisor isolation to distributed Ray workloads with fast startup and low overhead.

GKE Rapid channel: 1.36.4-gke.1391000 is new default

Rapid channel default moves to 1.36.4-gke.1391000; adds 1.34.11, 1.35.8, 1.36.4 and alpha 1.38.0 preview versions.

GKE ambient networking enters Preview with sidecarless service mesh

Preview allows allowlisted clusters to use node-level proxy for simplified L4 service mesh without sidecars.

Networking for AI inference model serving - GKE only and for all other backends

Reference architectures for exposing AI inference models behind a secure, private entry point using Cloud Load Balancer, Inference Gateway, Apigee, and Model Armor for governance, safety, and routing.

Now in Public Preview: AI-assisted EKS-to-GKE migrations

Open-source GKE Agentic Migration plugin automates EKS-to-GKE moves with deterministic guardrails and offline validation.

k8s-aibom open-source tool auto-generates AI SBOMs on GKE

Lightweight controller discovers AI runtimes and produces CycloneDX ML-BOMs to reduce shadow AI risk.

Multi-cluster GKE Inference Gateway for global GPU/TPU pooling

How-to shows routing traffic across regions with LLM-d router to treat worldwide accelerators as one pool.