The old model put a firewall at the boundary, treated the inside as trusted and the outside as hostile. In a cloud estate that stopped describing reality a while ago. Your services talk to each other and to Azure's own APIs over the public backbone, authenticated by tokens, so the check that matters happens on each call.
Managed identities instead of stored secrets
A credential in config or a pipeline variable can end up in a log, a repo or a screenshot. A managed identity has nothing to leak: Azure issues and rotates the credential and your code never sees it. In Terraform it's a few lines:
resource "azurerm_user_assigned_identity" "app" {
name = "app-identity"
resource_group_name = azurerm_resource_group.app.name
location = azurerm_resource_group.app.location
}
resource "azurerm_role_assignment" "app_reads_storage" {
scope = azurerm_storage_account.data.id
role_definition_name = "Storage Blob Data Reader"
principal_id = azurerm_user_assigned_identity.app.principal_id
}
The app authenticates as that identity and gets the Storage Blob Data Reader role on that one account.
Conditional Access, just-in-time admin and least privilege
Conditional Access decides who is signing in, from where, on what device and at what risk level, and what they must pass to get through. Requiring multifactor for admins and blocking legacy authentication both sit here, which makes it a bigger lever on your posture than any network ACL.
Privileged Identity Management activates a role for an hour when it's needed and lets it expire. A permanent Owner assignment gives an attacker standing access, whereas a stolen session with no standing roles has little to use.
For roles, a custom role you can describe in one sentence is easier to review than a pile of Owner grants whose approvals nobody remembers.
The full, runnable version is in the blog-examples repo.
Next: why your architect should be able to tell you what the design costs before anything ships.