Everything Azure Shipped This Week, Ranked

The week in one paragraph

Fourteen announcements, distributed unevenly: nine landed on Tuesday 1 September, two on Wednesday, one on Thursday, two on the Monday, and nothing at all on Friday. The Tuesday cluster is the week — four container items and three that together change what it costs to keep logs.

The item most likely to change an architecture is the one with the fewest words attached to it: a preview that puts Microsoft in the middle of your connection to another cloud provider.

Azure becomes the operator of your cross-cloud link

Azure Multicloud Interconnect entered public preview: a managed service providing private connectivity between Azure and supported cloud providers, with Amazon Web Services as the first supported provider.

Microsoft's framing of the problem is accurate and worth quoting, because it is the actual state of most large estates: as organisations distribute applications, data, AI services and business-critical workloads across multiple clouds, establishing and operating cloud-to-cloud connectivity becomes complex and operationally intensive. Today that connectivity is usually either a pair of ExpressRoute and Direct Connect circuits meeting at a colocation provider, or a set of IPsec tunnels somebody owns and nobody enjoys owning.

A managed service replaces the middle of that with a bill and an SLA. The architectural questions it raises are the ones to ask during preview rather than after: where the traffic physically traverses, what the failure domain looks like when one provider's side degrades, how routing and address-space overlap are handled, and whether the egress charges land differently than they do today. None of those are answered by a preview announcement, and all of them determine whether this is a simplification or a new dependency in the path between two clouds.

Worth noting where this sits in the week: it is the only announcement that changes what Azure is willing to operate on your behalf, rather than what it does inside its own boundary.

A container day

Four announcements on 1 September point at containers, and read together they are a maturity statement rather than four features.

Windows Server 2025 on AKS reached general availability, with Stable ABI, Generation 2 VM as the default, containerd 2.0 and FIPS enablement by default. Stable ABI is the one that matters operationally: it loosens the historically rigid coupling between Windows container image versions and host versions, which is the single biggest reason Windows node pools are painful to patch. Microsoft frames the release around support timelines for older Windows Server versions approaching retirement, which is the polite way of saying this is now a migration deadline rather than an option.

Artifact streaming on AKS went GA, letting workloads scale without fully waiting for images to be pulled into the cluster. For anyone running large images — and AI workloads routinely run very large ones — image pull time is the dominant term in cold-start latency, so this is a scaling-behaviour change dressed as a registry feature.

Confidential VMs for Azure Linux reached GA for AKS node pools, aimed at migrating highly sensitive container workloads without any code refactoring. Nodes in such a pool use a customised Azure Linux 3 image configured for CVM.

And Defender for Cloud extended its Serverless Containers Posture experience to Azure Container Apps environments, assessing identity, networking and container or image configuration alongside the rest of the container estate. The pattern across all four: the container platform is being levelled up at the edges — Windows, security posture, confidential workloads, start-up time — rather than in the core.

The cheap-logs trio

Three separate announcements on the same day add up to a single, genuinely useful change in the economics of observability.

The Auxiliary table plan in Azure Monitor Logs is the cost-effective tier for high-volume, verbose logs kept for compliance and auditing but rarely queried. Two things changed. It now supports a subset of standard Azure tables rather than only DCR-based custom _CL tables — meaning low-value data can stay on its original table, keeping its name, schema and existing queries, instead of being reshaped through a custom pipeline. And plan switching arrived, so a table's plan is now a decision you can revise.

Separately, the plan reached general availability in the sovereign clouds, Azure Government (Fairfax) and China.

The third piece closes the obvious gap. Moving logs to a cheap tier is only attractive if the tooling can still read them, and Azure Copilot Observability Agent now covers Log Analytics data in Basic and Auxiliary table plans during interactive analysis and deep investigations. Microsoft names the exact telemetry this is for: container stdout and stderr, audit trails, and node syslog — the three highest-volume, lowest-value-per-line streams in a Kubernetes estate.

Taken together this is the first time the cheap tier has been a default-reasonable choice rather than a trade against investigability. The work it implies is unglamorous and worth doing: go through the tables, identify which are queried and which are merely retained, and move the second group.

Networking: two policies and a preserved source IP

Azure Firewall auto-learn SNAT routes reached GA. It periodically learns registered and private destination prefixes and applies them as No-SNAT ranges, preserving original source IPs. Anyone who has maintained that list by hand knows why this matters: the list drifts every time a new address space is added, and the symptom of drift is that traffic to an internal destination arrives with the firewall's address instead of the client's, breaking logging, allow-lists and forensics all at once.

Azure Front Door WAF gained profile and route level policy associations in public preview. Policies can now be attached at the profile, domain or route scope — a broad baseline across the profile, with targeted policies for the applications, domains or routes needing different controls. That is the same layered-scope shape Azure Policy and RBAC use, arriving in the WAF.

Azure Virtual Network Manager IPAM reached GA in more regions, including US Gov Virginia, Texas and Arizona, and China North 3 and East 3. Centralised IP planning matters most where address space is most constrained and multi-cloud overlap most likely — which, given the Multicloud Interconnect preview above, is a pairing worth noticing.

On the data side, Azure Database for MySQL Flexible Server gained a reader endpoint in preview: one endpoint that load-balances read-only traffic across replicas, on a service supporting up to 10 read replicas. It removes the read-replica connection logic applications have been carrying themselves.

A failure mode gets smaller

Per-disk resiliency for Azure VMs entered public preview in select regions, and it is a good example of a platform default being made adjustable rather than replaced.

The existing behaviour stays the default: when a VM loses access to an attached managed disk for an extended period, Azure shuts the VM down and powers it back on once disk connectivity is restored, with no customer action required. That is the right default, because most applications cannot cope with a disk vanishing underneath them.

With per-disk resiliency enabled, Azure instead detaches and takes offline only the affected disk, while the VM and its remaining disks keep running. The eligibility test is stated in the announcement and is worth taking literally: this is for applications that can tolerate the temporary loss of an individual data disk. A database whose data files are on that disk cannot. A node whose scratch or cache volume is on it very likely can, and would rather stay up.

Security advisories

Six Azure advisories, all published on Wednesday 3 September, and the distribution is notable — five of the six are elevation of privilege, and four of those sit in identity or AI services rather than infrastructure:

• Identity — CVE-2026-62916 Microsoft Entra ID elevation of privilege, and CVE-2026-83711 Azure Active Directory B2C elevation of privilege. An elevation advisory against the identity plane is the one to read first.
• AI and data platform — CVE-2026-70352 Azure AI Language, CVE-2026-80098 Copilot Studio, and CVE-2026-70178 Microsoft Fabric, all elevation of privilege.
• Data — CVE-2026-69857 Azure Cosmos DB spoofing.

These are platform-side advisories rather than agents running on your servers, so the remediation is Microsoft's. The useful response is to check whether any of the affected services are in scope for your estate and to read the individual entries for customer action, which occasionally exists.

Beyond the announcements

Four items from Microsoft's blogs worth more than the announcement feed carried:

• Your private endpoint does not cover agent egress — locking down Azure AI Foundry in both directions. The best technical piece of the week: private endpoints secure inbound, and an agent's outbound calls are a separate problem that most designs have not addressed.
• Keep the evidence: low-cost logs with Azure Copilot Observability Agent — the practitioner version of the cheap-logs trio above.
• Automating compliance scope enforcement with Azure Policy — scope enforcement rather than rule authoring, which is the half of policy work that usually gets improvised.
• What's new in Azure Firewall — the consolidated view, published the same week as the SNAT change.

Also this week: instant revocation of service principal bearer tokens with CAE, ten posts from the Foundry blog, and releases from AKS, the Azure CLI, Bicep and the Terraform AzureRM provider.

What I would act on

1. Audit your Log Analytics tables against the new Auxiliary support. Standard tables and plan switching remove the two reasons not to use the cheap tier, and the agent now reads it. Container stdout, audit trails and node syslog are the obvious first three.
2. Plan the Windows Server 2025 move on AKS. Microsoft is explicitly citing retirement timelines for older versions, and Stable ABI makes this migration materially easier than the last one.
3. Stop maintaining the No-SNAT list by hand. Auto-learn is GA. A stale list produces the kind of source-IP loss that is discovered during an investigation, which is the worst time.
4. Read the Multicloud Interconnect preview before someone builds around it. A managed private path to AWS is attractive enough to be adopted before the failure domains are understood. Preview is the right time to ask those questions, not after.
5. Check which of your workloads could take per-disk resiliency — specifically ones with a cache or scratch data disk whose loss should not restart the VM. It is preview and region-limited, but the eligibility question is worth answering now.

Complete inventory — all 14

Every announcement published to Azure Updates between Monday 31 August and Friday 4 September 2026.

Official Azure references

• Azure Multicloud Interconnect
• Windows Server 2025 on AKS
• Artifact streaming on AKS
• Confidential VMs for Azure Linux
• Auxiliary Logs Plan support for Azure tables and plan switching
• Azure Copilot Observability Agent supports Basic and Auxiliary table plans
• Azure Firewall auto-learn SNAT routes
• Azure Front Door profile and route level WAF policies
• Per-disk resiliency for Azure VMs
• Reader endpoint for Azure Database for MySQL