This week in Azure
Small week by volume, strong week by content. Standard service endpoint is the one that will change designs, and the new protected table for generative AI telemetry is the one nobody will talk about and everybody in a regulated shop needs.
Networking carries the rest: IPv6 on VPN Gateway goes GA, DDoS Protection gets per-resource thresholds, and AKS can finally encrypt NFS traffic to Azure Files.
Networking
Standard service endpoint (preview)
Standard service endpoint is in public preview, and it is the fix for everything that made classic service endpoints awkward at scale. It joins the Private Link family rather than the old service-endpoint model.
The mechanism: instead of binding a subnet to a service, you associate a single public IP or prefix with multiple subnets or virtual networks, across subscriptions, within the same tenant and region. Public IPs become the network identifier. Combine it with network security perimeter and you can draw a boundary around Storage, SQL, Cosmos DB and Key Vault and permit only approved networks and workloads through.
Two constraints before you plan around it. Coverage today is Storage, SQL Database, Cosmos DB and Key Vault, expanding over time. And unlike classic service endpoints, which are free, this is a paid service — so the comparison against private endpoints is now a cost comparison too, not just an architectural one.
Still, this is the answer to the two complaints classic service endpoints always attracted: they did not scale across subscriptions, and they gave you no perimeter concept. If you have been reaching for private endpoints purely because service endpoints could not express your topology, re-evaluate. Preview, so prototype it — but design against it now, because it changes the shape of the answer.
IPv6 support for Azure VPN Gateway (GA)
Azure VPN Gateway now carries IPv6 inner traffic through the tunnel in a dual-stack configuration, supported on all production SKU gateways using standard public IPs.
Both directions are covered. Site-to-site carries IPv6 inside the IPsec/IKE tunnel for branch and on-premises connectivity. Point-to-site supports IPv6 in the client address pool and routed traffic configuration. It works on new and existing deployments.
You add an IPv6 range to the existing gateway to make it dual-stack. Two protocol exclusions to check first: IKEv1 is not supported for site-to-site, and SSTP is not supported for point-to-site. Everything else works.
Dual-stack, not IPv6-only, so the same gateway configuration serves both address families. If your IPv6 rollout stalled at the VPN boundary, that blocker is gone.
Azure DDoS Protection custom policy (preview)
DDoS Protection custom policy adds per-resource control over mitigation thresholds for protected Standard Load Balancer frontend IP configurations. You set fixed inbound detection thresholds for TCP, UDP and TCP SYN traffic, anywhere from 50,000 to 2,000,000 packets per second.
The point is predictable traffic that looks like an attack. Product launches, seasonal peaks, gaming events, sustained high-volume workloads. Adaptive auto-tuning learns your baseline, but it learns it from history, and a planned tenfold spike has no history.
Configure a threshold per protocol and Azure uses your number instead of auto-tuning for that protocol only; the rest stay adaptive. Sensible separation. Available through portal, CLI, ARM templates and REST during preview, inbound traffic on Standard Load Balancer frontends.
Kubernetes and storage
Encryption in transit for Azure Files NFS on AKS (GA)
AKS now supports encryption in transit for Azure Files NFS v4.1 volumes through the Azure File CSI driver. Traffic between AKS workloads and the NFS share is encrypted with TLS, using the AZNFS mount helper and Stunnel underneath.
You enable it in the storage class. No application changes. Available in every region that supports Azure Files Premium SSD.
NFS-without-encryption was a recurring finding in every compliance review that touched AKS persistent storage, and the workaround was usually “move to SMB” or “accept the risk”. Neither was a good answer. Turn this on when you next touch your storage classes.
Compute
Azure Functions: Python 3.14 GA, PowerShell 7.6 preview
Python 3.14 is generally available on Azure Functions for Linux plans. PowerShell 7.6 is in preview across Functions plans. Both develop locally and deploy normally.
Nothing clever here, but language-version currency is a security posture item, not a developer-convenience item. If you are running an out-of-support runtime, the upgrade path is now available for both.
Database
Defender security assessments for PostgreSQL Flexible Server (GA)
Microsoft Defender CSPM assessments for Azure Database for PostgreSQL Flexible Server are generally available. They evaluate your servers continuously against PostgreSQL-specific best practices and surface misconfigurations with remediation recommendations, visible on the resource blade in the portal.
If Defender CSPM is already enabled for the server, there is no additional setup. An initial set of built-in assessments ships now, more planned. Free posture signal on a service that usually gets audited manually — enable it and read what it says.
AI and observability
A protected table for generative AI telemetry (preview)
This is the quiet one that matters. Application Insights now stores generative AI content in a dedicated GenAIContent table, surfacing as AppGenAIContent in Log Analytics, so you can apply access controls to it separately from the rest of your telemetry.
Think about what actually lands in AI telemetry: prompts, responses, tool calls. In a healthcare or financial workload that is PII and PHI sitting in the same workspace your on-call engineer queries at three in the morning. Until now the choice was to log it and over-share, or not log it and lose the audit trail.
The control is built into Log Analytics at the platform layer, not bolted onto Foundry, so you can configure any table to require a more restrictive role for data access. Application Insights powers observability in Microsoft Foundry, so this applies whichever surface you use.
If you run AI workloads under a regulatory obligation, this is the week’s most consequential update. Not because it is difficult, but because it removes the excuse for not separating those permissions.
Two new MAI models
Microsoft shipped two models pointed at opposite ends of the same trade-off.
MAI-Image-2.5 Pro targets maximum visual fidelity and creative control, with strong object consistency across a series of images — the thing that usually falls apart when you generate a set rather than a single picture.
MAI-Voice-2 Flash goes the other way: low latency above all, across 15 languages. Call centre agents, voice assistants, anything interactive where milliseconds are the requirement rather than fidelity.
Both are worth reading as a statement about model selection rather than as two product launches. The idea that you pick one large model and route everything through it stopped making sense a while ago. You choose per task, on complexity, context, tool calling, reasoning depth and latency — the same way you would not put your most expensive engineer on every ticket.
Final thoughts
Two things to act on now.
Standard service endpoint changes how you connect IaaS to PaaS. Do not deploy it into production yet, but stop designing around the classic service-endpoint limitations, because they are about to stop being limitations. Map which of your private endpoints exist only because service endpoints could not scale.
The GenAIContent table takes an afternoon. Split the access role for AI telemetry away from general Log Analytics readers, and do it before someone asks who can read the prompts. Recovery from that question is much cheaper when the answer is already “a named role, and here is the list”.
Everything else is incremental. Encrypt the NFS traffic when you next touch your storage classes, and get the PostgreSQL assessments switched on since they cost you nothing.
Related deep dives
- Building a Governed AI Platform on Microsoft Foundry — the audit design the GenAIContent table plugs straight into
- Azure Cloud Foundation: The Boring Stuff That Matters — where service endpoints and perimeters get decided, once, for everyone
Sources
- John Savill, “Azure Update 24th July 2026,” YouTube, https://www.youtube.com/watch?v=fv-2mxaU8sw
- “Standard service endpoint overview,” Microsoft Learn, https://learn.microsoft.com/azure/private-link/service-endpoint-standard-overview
- “IPv6 configuration for Azure VPN Gateway,” Microsoft Learn, https://learn.microsoft.com/azure/vpn-gateway/ipv6-configuration
- “Azure DDoS Protection custom policy overview,” Microsoft Learn, https://learn.microsoft.com/en-us/azure/ddos-protection/ddos-custom-policy-overview
- “Encryption in transit for NFS shares,” Microsoft Learn, https://learn.microsoft.com/en-us/azure/storage/files/encryption-in-transit-for-nfs-shares
- “Create and manage volumes with Azure Files in AKS,” Microsoft Learn, https://learn.microsoft.com/en-us/azure/aks/create-volume-azure-files
- “Configure protected tables in Azure Monitor Logs,” Microsoft Learn, https://learn.microsoft.com/azure/azure-monitor/logs/protected-tables-configure
- “Generative AI telemetry data model,” Microsoft Learn, https://learn.microsoft.com/azure/azure-monitor/app/data-model-complete
- “Language runtime support in Azure Functions,” Microsoft Learn, https://learn.microsoft.com/en-us/azure/azure-functions/supported-languages