The short version
Microsoft runs a demonstrably secure platform — and that fact is routinely mistaken for a guarantee that the workloads on it are secure. They are not. Under the shared responsibility model, Microsoft secures the hypervisor, the physical estate, the control plane and the service fabric. Everything above that line — network reachability, identity scope, image provenance, egress control, key custody — is the customer's to configure, and Azure ships those defaults tuned for onboarding speed, not adversary resistance.
This paper is a control-by-control reference for closing that gap across two planes: the network and perimeter, and compute and workload. They only work as a pair. A hard perimeter around a soft interior fails; so does a hardened machine that still calls out to an attacker-controlled endpoint. The gap between an average Azure estate and a hardened one is, with few exceptions, a matter of configuration discipline and enforcement architecture rather than budget or capability.
What's inside
- Part I — Network and perimeter hardening. Topology and segmentation, network and application security groups, egress architecture, Azure Firewall, Private Link and private DNS, web application firewalling, DDoS protection, and network-plane telemetry.
- Part II — Compute and workload hardening. VM platform integrity, the image supply chain, encryption at rest, administrative access paths, patching, Azure Kubernetes Service, App Service and Functions, and secrets and key management.
- Part III — Enforcement. Azure Policy as the control plane, and a rollout pattern that does not cause outages.
- Part IV — Azure Government deltas. Where the endpoints, compliance posture, and operational realities diverge from Azure Commercial.
- A threat model mapping the canonical attack path to a control at each stage, ten design principles, a phased implementation roadmap, and a condensed hardening checklist for design review and assessment.
Key findings
- Egress is the control you are most likely missing. Inbound filtering is well understood and usually present; outbound filtering is frequently absent — and it is the path command-and-control, token exfiltration, and data theft all traverse.
- Private Link is a DNS project that happens to involve networking. Most failed private-endpoint deployments fail at name resolution, not connectivity, and a private endpoint that coexists with an enabled public data plane provides the appearance of isolation without the substance.
- Workload identity has replaced secrets, but most estates have not finished migrating. Long-lived service principal secrets and connection strings persist in application settings, pipeline variables, and Kubernetes secret objects because nobody owns the migration.
- Policy, not documentation, is the enforcement boundary. A hardening standard that lives on a wiki page decays within a quarter; the same standard expressed as Azure Policy survives staff turnover, holds under deployment pressure, and generates its own audit evidence.
- Defaults are not a baseline. Azure's shift to private-by-default subnets for virtual networks created after 31 March 2026, and AKS clusters that still ship with local admin accounts enabled, are two of several defaults the paper shows working against a hardened estate rather than for it.
Who it's for
Cloud platform and infrastructure engineers responsible for landing zone design and implementation; security architects reviewing an Azure design against a control framework; and technology leaders who need to understand what "hardened" concretely means before funding it. Sections are written to be read independently — a network engineer can start at Part I, a Kubernetes team at the AKS section, without loss of context.