Everything Azure Shipped This Week, Ranked

The week in one paragraph

Fourteen announcements, and three of them landed on the same product on the same day. Azure Bastion picked up expiring shareable links, IPv6 in preview, and — the one that changes a design — a tunnel to the AKS API server. Azure SRE Agent got network integration, reports and a trial, which together move it from a demo to something that could plausibly sit inside a production network boundary. Underneath both, a small storage announcement removed one of the last routine reasons an AKS cluster holds a storage account key.

No Azure security advisories were published in the window. One retirement carries a date and a required action, and it is the kind that will be ignored until the quarter it bites.

Bastion stops being a VM console

The headline item is Azure Bastion integration with AKS, now generally available. It establishes a secure tunnel from a local machine, through Bastion, to an AKS cluster API server. Microsoft's framing is worth reading precisely: it lets customers use standard Kubernetes tools while avoiding public exposure of private cluster endpoints, and it reduces the need to deploy and maintain separate jump boxes, VPN servers, or additional access agents.

Anyone who has run a private AKS cluster knows what that sentence is describing. The cluster is private, which is correct, and then a jump box appears next to it so that humans and pipelines can reach the API server — a virtual machine with credentials, an operating system, a patch cycle and a login history, existing solely as a doorway. It is the least interesting machine in the estate and one of the most sensitive. Bastion is a managed service that already terminates sessions, already logs, and already sits in the hub. Moving the doorway onto it removes a server rather than adding one, which is the rarer kind of platform improvement.

Alongside it, shareable link expiration reached GA: an administrator can set the date and time a shareable link stops working, after which it can no longer be used to connect to the target resource. Shareable links have always been the convenient and slightly uncomfortable part of Bastion — access handed to someone outside the tenant, revocable only by remembering to revoke it. An expiry turns that from a standing grant into a bounded one, which is the same argument this series made about policy exemptions yesterday: a waiver with an end date is a decision, and a waiver without one is an omission.

IPv6 dual-stack support entered public preview, with two constraints stated plainly and worth carrying into any plan. IPv6 covers connections between the user and Bastion; connections from Bastion to target virtual machines continue to use IPv4. And existing IPv4-only Bastion deployments cannot currently be converted to dual stack — so this is a new-deployment feature, and a dual-stack estate needs a new Bastion rather than an upgraded one.

An operations agent with a network address

Azure SRE Agent VNet Integration reached general availability, and it is the announcement in this week's set with the largest gap between how routine it sounds and what it actually grants. VNet integration lets the agent operate within existing network controls — network security groups, private DNS, firewall policies — and, in Microsoft's words, securely access private resources, including services behind private endpoints, during incident investigation and remediation without requiring changes to the network boundary.

Read that as an access-control statement rather than a networking one. An automated agent that can reach services behind private endpoints, investigate them, and remediate is a principal with production reach. The networking half is the part that gets announced; the half that needs designing is the identity, the scope and the audit trail. That is not a criticism of the feature — the alternative, an agent that only works when the boundary is opened for it, is worse — but the phrase "without requiring changes to your network boundary" describes what you do not have to do, not what you should not check.

The other two SRE Agent items make it likelier that teams will try it. Live Reports entered public preview, generating operational views from agent conversations and refreshing them on a daily, weekly or monthly cadence. And a 30-day trial is now available for new customers with no baseline always-on charges, paying only for Azure Agent Units consumed when agents perform work. A consumption-only trial is the right shape for something whose value is impossible to predict from documentation, and the pricing model — work performed, not capacity reserved — is the one detail worth confirming before anyone points it at a busy environment.

One more place storage keys were still living

Workload identity support for the Azure Files CSI driver (SMB) reached general availability in AKS. The progression Microsoft describes in a single sentence is the whole story: managed identity support previously let workloads mount Azure Files without using storage account keys; workload identity now takes it to pod-level authentication, so application pod access can be scoped to exactly the data it needs rather than granted at the node level.

Node-level access is the compromise almost every AKS cluster with file shares has been living with. Every pod scheduled on a node inherits what that node can mount, so the blast radius of a compromised container is the union of every share any workload on that node uses. Pod-level identity collapses that to the pod's own scope, and it is available in all Azure regions that support Azure Files and AKS — no staged rollout to wait out.

This is a small announcement with a large cleanup behind it. If a cluster still holds storage account keys in secrets for SMB mounts, the reason it does is now historical.

Two different answers to a runtime running out

Two announcements this week address the same problem from opposite ends, and reading them together is more useful than reading either alone.

The first is a hard stop. Support for Node 22 LTS on App Service ends on 30 April 2027. Apps hosted on App Service will continue to run — that is the sentence that causes the deadline to be missed — but security updates will no longer be available and Microsoft will no longer provide customer service for Node 22 LTS. The required action is stated: upgrade to Node 24 LTS before that date. An application that keeps serving traffic on an unpatched runtime is not a working application, it is an unresolved incident with a delay on it.

The second is the commercial alternative. Extended Support for Azure Database for PostgreSQL Flexible Server was announced, giving access to critical security updates, critical bug fixes and technical support for eligible versions after community support ends. That is a paid deferral rather than a fix, and it is the honest shape for a database: major-version upgrades of PostgreSQL are project work with a maintenance window, and the ability to buy time on a supported footing is worth more than the version number.

The pattern to take from the pair: platform runtimes get a date and an upgrade path, and data engines get an extended-support option. Plan the first as an engineering task and the second as a budget line.

Compute and container networking

248 and 372 vCPU sizes for the Dl/D/E v7 series reached general availability on Intel Xeon 6, with up to 20% better compute performance than the prior-generation Intel-based v6 sizes. The numbers that matter for placement decisions: up to 372 vCPUs and 2.8 TiB of memory, and up to 400 Gbps networking bandwidth on the 372 vCPU Esv7/Edsv7 size. Microsoft positions it for larger in-memory databases and for agentic workloads with larger context windows, and notes the reduction in cross-node network hops — which is the real argument for scaling up rather than out on latency-sensitive systems.

Worth pairing with last week's vCore customisation announcement if you run per-core licensed software: a 372 vCPU host is a large licence bill unless the licensable core count is set deliberately.

eBPF host routing in Advanced Container Networking Services for AKS reached GA, moving packet forwarding and routing decisions into the Linux kernel with eBPF and removing the overhead of iptables-based processing. The operational detail is the one to note: enabling it on an existing cluster triggers a controlled node rollout that respects existing network connections. That is a rolling change to the data path of a live cluster, and it should be treated as such even though it arrives as a feature flag.

Custom block response code and body for Application Gateway WAF is now GA, bringing it in line with WAF on Front Door. By default a blocked request returns a 403 with "The request is blocked"; a custom status code and message can now be set instead. It is a policy-level setting, so every blocked request in that policy gets the same response — useful for hiding that a WAF is present, and useful for giving legitimate callers something actionable, but not both at once.

Finally, Azure VM Image Builder became generally available in Azure Government, China North 3, Azure Government Secret and Azure Government Top Secret. For teams running the same platform across global and sovereign clouds, the value is that the image pipeline definitions stop being environment-specific.

Security advisories

No Azure advisories were published to the MSRC Update Guide in this window. Last week's Arc and Connected Machine Agent advisories are the ones still worth confirming as patched, since those agents run on servers you own and wait for your schedule rather than Microsoft's.

Beyond the announcements

Four items from Microsoft's blogs that carry more than the announcement feed did:

• Instant revocation of service principal bearer tokens with CAE — the gap between disabling a service principal and its tokens actually failing has always been the awkward part of workload identity incident response. Continuous access evaluation closing it for service principals is the most consequential identity item of the week.
• Advertised gateway prefixes in Azure — the follow-up detail to last week's summarised-prefix GA, which is the piece anyone planning summarisation actually needs.
• The Standard HDD I/O unit size update — a change to how I/O is counted is a change to a bill, and this one arrives without a pricing announcement attached.
• Zonal redundancy in API Management Standard v2 — zone resilience arriving on a mid-tier SKU rather than only the premium one, which is the same pattern as last week's SQL Managed Instance preview.

Also this week: confidential live migration, the retirement of Microsoft HPC Pack, regional host pools for Azure Virtual Desktop in preview, simplified regional limits for SQL Managed Instance, and a Terraform AzureRM v5.3.0 release.

What I would act on

1. Count your jump boxes. Bastion-to-AKS is GA, and every jump box that exists only to reach a private API server is now a server you are choosing to keep patched. Removing one is a security improvement and an operational saving in the same change.
2. Put an expiry on every Bastion shareable link, and go back over the existing ones. The feature is new; the links handed out before it are still open.
3. Diarise 30 April 2027 for Node 22 on App Service. The apps keep running, which is exactly why this gets missed. Upgrading to Node 24 is a scheduled task now and an emergency in eighteen months.
4. Replace node-level Azure Files access with pod-level workload identity. It is GA in every region that supports both services, and it retires the storage account keys that a cluster has probably been carrying since it was built.
5. If you are evaluating SRE Agent, design the identity before enabling VNet integration. Reaching private endpoints to investigate and remediate is real production access, and the trial makes it easy to grant that quickly.

Complete inventory — all 14

Every announcement published to Azure Updates between Monday 24 and Friday 28 August 2026.

Official Azure references

• Connect to AKS clusters using Azure Bastion
• Azure Bastion shareable link expiration
• IPv6 dual-stack support for Azure Bastion
• Azure SRE Agent VNet Integration
• Workload identity support for Azure Files CSI driver (SMB)
• Retirement: Support for Node 22 LTS ends on April 30, 2027
• Extended Support for Azure Database for PostgreSQL Flexible Server
• 248 and 372 vCPU sizes for D/E v7 series VMs
• eBPF host routing in Advanced Container Networking Services for AKS
• Custom block response code and body for Application Gateway WAF