What Microsoft actually announced
Microsoft describes the service as an AI-driven, preemptive security service that protects Azure resources from known and emerging vulnerabilities, including zero-day threats. The pitch, from Microsoft's Azure Networking blog, is about timing. A critical flaw is disclosed, security teams find the affected workloads, application owners test the patch, and attackers move faster than any of that.
The service is meant to cover that gap. It combines vulnerability intelligence, details of what each workload is running, and enforcement in two places: the network and the workload itself.
The first point to understand is that this is not a patching tool. Microsoft's documentation says the service doesn't remediate the vulnerability or replace patching. Microsoft's blog says the same: customers keep validating and deploying fixes through their normal processes. Once the patch is in and remediation is confirmed, the temporary mitigations can be rolled back.
Scope: Linux first, Windows later
The launch scope is narrow. Microsoft says it is starting with Linux virtual machines and Virtual Machine Scale Sets on Azure. The company also says it plans to extend protection to other workload types, including Windows-based workloads. No date has been given for Windows support.
For Windows-centric shops, that means this preview is a watch-and-wait item for now. Mixed estates with Linux web front ends, such as PHP or HTTP/2 servers sitting beside Windows back ends, are the more likely early testers.
How it works
Microsoft describes three layers:
- Vulnerability knowledge. A graph links vulnerabilities to affected components, exploit prerequisites, vulnerable behaviors and possible mitigations.
- Customer workload context. A map of your Azure estate is built from telemetry: running processes, installed modules, operating systems, open ports and similar details. Microsoft's reasoning is that having a vulnerable package installed is only part of the risk. What matters more is whether it is running and whether it can be reached.
- Orchestration. An AI core combines the first two layers. It then picks network controls, workload runtime controls, or both.
Network and workload enforcement
On the network side, Microsoft says multiple defenses, including inline intelligence, can interrupt exploit-related traffic and limit the paths to vulnerable workloads.
On the workload, enforcement uses eBPF through Tetragon. Microsoft says the AI core synthesizes Tetragon policies for each workload's software, configuration and exploit conditions. A "neuron extension" installed on each onboarded workload does two jobs. It sends telemetry back to the service, and it deploys the mitigations the AI core instructs.
AI picks the control, rules enforce it
Microsoft stresses a design choice: AI decides what protection is needed, but once a policy is deployed it runs as fixed rules. Microsoft says the deployed policies enforce locally and don't need a new generative-AI decision for every matching event.
That should make enforcement more predictable. It does not remove the need to check that the AI-generated policy is correct in the first place. Microsoft hasn't published how that correctness is validated.
Preview limits and pricing
Microsoft's documentation lists several limits that matter for planning:
- You can't mix individual VMs and VM scale sets in the same Adaptive Shield.
- You can create up to two Adaptive Shields per subscription. Each shield can protect up to 40 individual VMs, or one or more scale sets with up to 40 instances combined.
- During preview, Microsoft says to deploy it in test environments rather than production.
- The documentation also lists a severity threshold setting for each shield (Low, Medium, High or Critical). It controls how sensitive threat detection is.
On cost, pricing is based on the number of shields and the number of protected resources in each, and during preview you pay 50% of the price. Microsoft's product page adds that there is no separate zero-day surcharge. The number of vulnerabilities or virtual patches is not a billing factor. I don't have the actual per-unit rates, so check Azure pricing for your region.
Microsoft's test claims
Microsoft reports two examples from its own testing:
- HTTP/2 Bomb (CVE-2026-49975). This is a resource-exhaustion attack that uses header compression amplification and stalled streams. Microsoft says the service mitigated it in testing.
- PHP2Shell (CVE-2026-109618 and CVE-2026-109619). These are PHP runtime memory-handling flaws that can allow unauthenticated remote code execution under exploitable conditions. Microsoft again reports successful mitigation in testing.
These are vendor-reported results. Nothing in the material I reviewed is an independent benchmark. Microsoft's own caveat on PHP2Shell is worth repeating: interim measures don't cover every exploitation route, and installing the fixed software is still essential.
Practical guidance for admins
- Treat it as a test-environment trial. Microsoft's own guidance for the preview is non-production use.
- Check compatibility with your existing tools. Microsoft says it can complement other security tools, but compatibility depends on the specific solution and configuration. Test that interaction before you rely on it.
- Keep patching. Azure Update Manager installs and schedules real updates. This service only blocks exploitation attempts.
- Watch for side effects. Runtime policies that block process, file or network behavior could affect legitimate applications. Microsoft's material doesn't detail how false positives are handled. Verify that in a lab.
- Plan the rollback. Remove mitigations only after the patch is deployed and confirmed.
- Check prerequisites and regions. The documentation lists prerequisites and supported regions. I didn't get the full details, including supported Linux distributions or required permissions, so confirm them before you start.
Bottom line
Virtual patching isn't new. Network and web application firewalls have offered versions of it for years. What's different here is the claim that AI-built mitigations can be tailored to each workload and enforced at both the network and the host. That is a bigger promise, and so far the evidence is Microsoft's own testing.
The preview limits are modest: two shields per subscription, 40 VMs per shield, Linux only. That is enough for a pilot and not enough for a fleet. Windows admins should keep an eye on the promised Windows workload support.
References
- (In preview) Public Preview: Azure Adaptive Virtual Patching Azure Updates · 2026-11-17T20:29:13Z
- Adaptive Virtual Patching | Microsoft Azure azure.microsoft.com
- Azure Adaptive Virtual Patching overview learn.microsoft.com