Azure Security Baseline: Hardening Subscriptions, Identities, and Networks

Azure breaches almost always start at Entra ID (formerly Azure AD), not at a VM or storage account. Get identity right and the rest gets dramatically easier:

• Conditional Access — require compliant device + phishing-resistant MFA for any privileged role.
• Privileged Identity Management (PIM) — make Global Admin and Owner roles just-in-time, with approval and time-bound activation.
• Workload identities — replace client secrets with Federated Identity Credentials (OIDC) wherever possible.
• Break-glass accounts — at least two, with FIDO2 keys stored physically, monitored aggressively.

Azure Policy is the structural prevention layer. Apply at the management group level so subscriptions inherit:

• Deny resources in unapproved regions.
• Deny public IPs on VMs unless explicitly tagged for it.
• Require diagnostic settings to send to a central Log Analytics workspace.
• Require encryption with customer-managed keys for sensitive data services.
• Require a tag set (cost-center, owner, environment) on every resource.

Use the Microsoft Cloud Security Benchmark initiative as a starting baseline rather than authoring from scratch.

The default "public endpoint" for storage accounts, SQL databases, and Key Vaults is convenient and dangerous. The pattern that holds up:

• Private Endpoints for every PaaS service that supports them — storage, SQL, Cosmos DB, Key Vault, App Configuration.
• Disable public network access at the resource level once private endpoints are in place.
• Service endpoints + NSG as a fallback where private endpoints aren't supported.
• Azure Firewall or third-party NVA at the hub for egress filtering.

Microsoft Defender for Cloud is Azure's CSPM/CWPP. It's good at:

• Surfacing common misconfigurations against the Microsoft baseline.
• Vulnerability assessment for VMs and containers.
• Adaptive application controls for VMs.

It's mediocre at:

• Custom detections — you'll layer Sentinel on top.
• Multi-cloud — works for AWS/GCP, but native tools are still better there.

Treat the Defender secure score as a tracking metric, not a goal in itself. A score of 80% with the right 20% missing is meaningfully different from 80% with random gaps.