Three AKS features went GA on Sep 1, 2026 — and together they tell you what Microsoft thinks platform teams need: faster startup for heavier images, modern Windows node images, and stronger isolation for sensitive Linux workloads.
The practical win is simple. Windows Server 2025 node pools arrive in AKS so teams running Windows containers get newer host OS support without shoehorning them onto older images; Artifact Streaming from Azure Container Registry lets container runtimes begin starting containers as layers stream rather than waiting for a full image download; and confidential VM node pools for Azure Linux provide hardware-backed runtime isolation for workloads that need enclave-level protections.
Artifact Streaming is the most operationally meaningful if you run large images or service rapid scale-ups. Instead of waiting for complete layer downloads, AKS nodes can start pulling and streaming layers on demand, which trims cold-start latency for jobs and short-lived containers. This isn't a feature you'll notice on tiny base images, but for 500 MB+ images and CI runners it can shave seconds to tens of seconds off start time — which adds up when autoscale events are frequent. Treat it like enabling a network-boot optimization: it makes cluster autoscaling and bursty CI far less painful.
Confidential VMs for Linux node pools are the hard security promise here: Intel TDX or AMD SEV-SNP-backed isolation for node VMs, depending on region and SKU. If you handle regulated data or multi-tenant secrets in a shared cluster, confidential node pools remove much of the 'we trust the host' assumption for Linux nodes. If you need the deep dive, I covered the GA in more detail earlier: AKS Confidential VMs GA: Azure Linux CVM Support, ContainerLogV2 & Telemetry Migration.
The upgrade story has one rough edge: Microsoft is retiring Azure Network Policy Manager (NPM) support for Windows nodes with a planned cutoff on Sep 30, 2026. That's a hard date — customers already onboarded must plan migration and new onboarding is disabled. If you run Windows node pools that rely on Azure NPM, this is not optional; it's a calendar event. Microsoft hasn't provided an automated in-place migration for every topology, so expect manual work: export your policies, validate an alternative (for example Calico where supported, or Azure's other enforcement modes), and test thoroughly. This retirement should have been coordinated earlier; teams still using Windows NPM are being forced to fix technical debt on a short runway.
AKS release-tracker data also showed Kubernetes 1.37 in preview in September 2026, with GA expected later; AKS will keep support for 1.36 into mid-2027. That schedule matters if you're juggling CRD compatibility, custom admission hooks, or node-agent constraints tied to specific kubelet behavior. If you haven't tested 1.36-to-1.37 behavior on your CNIs and runtime integrations, treat preview as your window to do that testing now. For context on 1.37 work, see our piece on the previewed kubelet-in-user-namespace and scale-to-zero HPA beta moves: Kubernetes 1.37: KubeletInUserNamespace (rootless) and HPA scale-to-zero Move to Beta.
Outside AKS: GitHub and Azure DevOps announced a Sep 4 batch of features: GitHub Advanced Security alert metrics integration, GitHub Copilot Code Review in public preview, Autofix AI credit billing, and a pipeline behavior change that stops automatically mounting the Docker socket by default for Linux container jobs. That last change is the right call; exposing /var/run/docker.sock has been a recurring foot-gun and pipeline images should opt into Docker-in-Docker explicitly if they need it.
Opinion: Azure is making the right bet here by hardening and optimizing infrastructure primitives rather than chasing surface-level AI feature announcements. Faster image startup and confidential node pools are the kind of operational improvements that reduce toil and risk. The flaky part is the cadence of deprecations: retiring Windows NPM with a one-month visible runway is bad process. Platform teams should assume Microsoft will nudge hard on deprecations going forward and bake migration capacity into quarterly plans.
Predict this: in twelve months AKS clusters will default to stricter runtime isolation and less permissive container mounts, and teams that haven't automated their policy migrations or image-size hygiene will be the ones with the loudest postmortems. If you run Windows workloads on AKS, treat the September deprecation date as the moment to stop deferring that migration.