kata-vm-isolation-v2, into public preview. It's meant to cut the cost of starting VM-isolated pods without giving up VM isolation. The Azure Updates listing is dated November 17, 2026. As summarized in the Azure Feeds mirror of that listing, the class is for running sandboxed pods with VM isolation. The aim is to reduce per-pod cold-start latency and memory overhead compared with classic per-pod Kata VMs. The listing's own teaser frames the goal as isolated pods that shouldn't force a tradeoff between security and startup performance.
What the announcement does and doesn't say
The "In preview" label has a specific meaning on Azure Updates. The page defines it as available to Azure customers for non-production use and testing, as opposed to "Launched," which means production-ready.
The material I could inspect does not give:
- supported AKS versions or regions
- the exact enablement steps
- performance numbers
- preview-specific terms or limits
Treat this as a signal to start planning a test, not as a deployment guide. I haven't seen Microsoft's documentation for the v2 class itself, so I can't confirm how it is enabled.
Background: how AKS pod sandboxing works today
The existing feature gives useful context for what v2 is changing.
- Pod sandboxing on AKS runs workloads in separate lightweight pod VMs. Each has its own guest kernel and is isolated from the host kernel and from other pod VMs.
- It is built on the open-source Kata Containers project.
- A pod selects the runtime with
runtimeClassName: kata-vm-isolation. Containerd then calls the Kata shim, which starts the Cloud Hypervisor VMM to create the pod VM. - Pod sandboxing is enabled per node pool with the
--workload-runtimesetting, and sandboxed pods can run alongside normal pods. - The documentation for the current class uses
KataVmIsolationas the node pool value, with--os-sku AzureLinux. Older preview docs used the namekata-mshv-vm-isolation. - Third-party examples from the Kubernetes agent-sandbox project say it needs Azure Linux nodes and a Gen2 VM size that supports nested virtualization. Treat that as a community observation. Check Microsoft's current documentation for hard requirements.
- Microsoft notes that switching to the Kata runtime class doesn't guarantee behavior identical to the standard runc runtime. That caution probably applies to v2 as well, since it is a different startup path.
The new class is a separate, additional name. Nothing in the announcement suggests the existing kata-vm-isolation class is being replaced.
Why a template-based start matters
Classic sandboxing boots a fresh VM for every pod, which costs time and memory. Microsoft says v2 instead creates pods from a pre-initialized sandbox VM template. Kata's own documentation describes VM templating as cloning new VMs from a pre-created template. Cloned VMs share the same initramfs, kernel and agent memory in read-only mode.
Kata's documentation also gives numbers for its upstream templating feature. In one example, 100 containers with 128 MB of guest memory each saved about 9 GB, roughly 72%. In another, ten containers started up to 38.68% faster. These are upstream results from an older setup, not measurements of Microsoft's AKS preview. Don't expect the same gains.
The security tradeoff to understand
Kata's upstream docs also list a drawback of VM templating. It cannot prevent cross-VM side-channel attacks of the kind described for Linux KSM, such as CVE-2015-2877. Kata's warning is that anyone concerned about that attack vector should avoid templating and KSM.
I can't say whether Microsoft's v2 implementation shares memory in the same way or mitigates this. The announcement doesn't address it. If you run mutually untrusting tenants, ask Microsoft or read the preview documentation before assuming the isolation is equivalent to the classic class. The reasoning here is mine, based on the upstream Kata documentation.
How to evaluate it
- Pick a non-production AKS cluster and wait for Microsoft's documentation on enabling v2.
- Check prerequisites: AKS version, region, node OS and VM size. Today's pod sandboxing needs Azure Linux nodes and a VM size with nested virtualization.
- Run the same workload under the existing
kata-vm-isolationclass and under v2. - Measure pod startup time, per-pod memory and pods per node.
- Test application compatibility, since sandboxed pods can behave differently from runc pods.
- Review the isolation model with your security team before any production use, and wait for general availability.
Bottom line
For teams that already use AKS pod sandboxing, this is a promising preview aimed at the main complaints about it: slow starts and memory overhead. Microsoft hasn't published figures, so the gains are unproven. Verify them with your own workloads.
References
- (In preview) Public Preview: Pod Sandboxing kata-vm-isolation-v2 runtime class Azure Updates · 2026-11-17T23:00:02Z
- Pod sandboxing in Azure Kubernetes Service (AKS) overview - Azure Kubernetes Service learn.microsoft.com
- Pod Sandboxing with Azure Kubernetes Service (AKS) - Azure Kubernetes Service learn.microsoft.com