Microsoft SC-500: Key Vault Access Design
Azure Key Vault is a security boundary, not simply a convenient place to put application passwords. An organization can create a vault, place secrets inside it, and still expose those assets through overbroad data-plane rights, permissive network settings, stale credentials, or incomplete monitoring. SC-500 expects engineers to reason across identities, policy, networking and threat protection at the same time. A practical design must say who can administer a vault, who can use each secret or key, where requests can originate, how credentials change, and how suspicious access will be detected and investigated.
On this page
- Separate vault administration from access to protected material
- Choose Azure RBAC without accidentally granting self-service privilege
- Combine Key Vault networking and DNS
- Treat secrets, keys and certificates as different lifecycles
- Replace static secrets with workload identity
- Enforce baselines with Azure Policy
- Investigate Defender for Key Vault alerts
- Verify the entire access journey before signing off
Separate vault administration from access to protected material
Azure Key Vault has two related but distinct security surfaces. The control plane concerns the vault resource: creating it, changing settings, and configuring platform properties. The data plane concerns objects stored in the vault: keys, secrets and certificates. Confusing these planes leads to a common review error. A principal may need to inspect vault configuration without needing to read a production secret. Another principal may need to retrieve a specific signing key without having authority to change the network policy around the vault. The permissions should follow those different responsibilities rather than a broad administrator label.
Protect a production vault by separating the identity that manages the resource from the application identity that retrieves a secret. For each principal, specify the actual operation, vault, object type, permitted environment and owner, then test the authorization boundary. A service may need to read one versioned secret but should not be able to change the vault firewall or enumerate unrelated keys. Check the configured RBAC model, access assignment scope, network rules and private access path before assuming a failed retrieval is a cryptography problem. After correction, validate both the permitted secret read and an intentionally rejected administrative action, and retain an audit event for each.
For deeper Microsoft-specific behavior, the Key Vault RBAC model distinguishes the two planes and their authorization mechanisms. Treat the current documentation and the deployed API version as the source of truth: the default access model for newly created vaults has changed over time, and a migration design must inspect existing vaults instead of assuming that new defaults were retroactively applied.
Choose Azure RBAC without accidentally granting self-service privilege
For data-plane access, Microsoft recommends Azure role-based access control over legacy vault access policies. New vaults created through API version 2026-02-01 or later default to Azure RBAC unless access policies are explicitly selected; existing vaults do not silently change their permission model merely because a newer API is used. An access-policy migration must inventory affected principals, preassign corresponding data-plane roles, and test access to avoid application lockout. A key security reason is separation of duties. With the legacy model, a contributor who can edit vault access policies can potentially grant themselves access to secrets. With Azure RBAC, assigning data-plane roles is governed through Azure role-assignment authority. That distinction does not remove the need to control Owners and User Access Administrators; it shifts the review to identities permitted to create or change role assignments.
An access design should start with the narrowest practical scope and purpose-built role. A service reading a secret should not automatically gain authority to create keys or manage certificates. A certificate-rotation operator should not also own unrelated production secrets. Resource-group and subscription scope can be tempting for operational convenience, but broader assignments amplify the impact of a compromised principal. The same distinction between role and scope appears in the existing Azure RBAC role-and-scope treatment and is directly relevant to a security review.
Use a review scenario to test the boundary: Can a deployment engineer who changes vault configuration also retrieve its application credentials? Can an emergency operator access production material outside the incident window? Can an application registration consent to permissions that its owner never intended? Document the authorized answer before granting access. For human elevation, apply Privileged Identity Management where appropriate, protect role-assignment authority and review standing privileges. A role name alone does not explain whether the overall access path is safe.
Reduce exposed paths with network controls and reliable DNS
Authorization is necessary but does not decide which network paths can reach the vault. A Key Vault that accepts requests from public networks has a different exposure profile from one reached through controlled private connectivity. Evaluate public network access, firewall rules and private endpoints in the context of the applications that must connect. A private endpoint provides a private IP path for the data-plane service, but it does not grant authorization. The requesting identity still needs the correct role and the client still needs working name resolution and route access.
DNS deserves its own acceptance test. Teams sometimes create a private endpoint and verify that it appears in the portal, but leave a workload resolving the public service endpoint or routing through the wrong network. Validate name resolution from the real client network, not just from an administrator workstation. Check split DNS, private zone links, hybrid resolvers and failover paths. A security control that breaks after region failover can create pressure to reopen public access as an emergency workaround.
The relationship with broader cloud administration matters: the production network is shaped by subnet routes, address spaces, DNS and policy inheritance. The Azure VNet and DNS material provides context for why apparently correct private access can fail when traffic and resolution are traced end to end.
Treat secrets, keys and certificates as different lifecycles
A secret is not equivalent to a cryptographic key or a certificate. An application secret might be a database credential or API token that needs rotation on a planned schedule. A cryptographic key may have usage restrictions and can be tied to encryption, signing or key wrapping. A certificate introduces an issuer, expiration date, renewal process, validation chain and consumers that must accept its updated form. A single generic rotation policy will not fit all three.
Inventory owners and consuming systems before changing versions. For secrets, establish how applications pick up a new value and how long the old value must coexist during a safe transition. For keys, distinguish rotating the active key version from re-encrypting protected data; not every consumer automatically moves when a new version exists. For certificates, track the real service endpoint and handshake after renewal. A rotation job that reports success while a workload still points to an expired object has not completed the protection cycle.
Plan recovery as well as normal rotation. Soft-delete and purge protection can reduce the impact of accidental or malicious deletion, but recovery procedures must be tested against the organization’s actual permissions and retention settings. Backup and restore options depend on object type and service constraints; never assume a vault backup substitutes for an end-to-end disaster-recovery plan. Operators should know who can restore an object, how downstream applications are reconfigured, and how the restoration is recorded.
Use workload identity to avoid turning secrets into configuration files
A managed identity can let a supported Azure workload request an access token without keeping a reusable service-principal password in configuration. That improves the credential-handling model, but it does not make the workload harmless. If the application is compromised, the attacker may exercise whatever data-plane permissions were assigned to that identity. The identity therefore needs a specific purpose and narrow role assignment, and it must not be casually shared across unrelated applications.
Prefer using the platform identity at the point where a workload retrieves its secret or key. Keep secret values out of deployment logs, build artifacts, browser-visible configuration and source repositories. Verify that diagnostic traces do not accidentally reveal material. An application should also degrade predictably if the vault is temporarily unreachable; repeated blind retries can worsen an outage and increase noise. For application-side context, secrets management for Azure AI apps discusses the same trust boundary in AI application deployments.
Enforce the baseline through Azure Policy and documented exceptions
A consistent vault posture cannot rely solely on the memory of the engineer who deployed it. Azure Policy can assess or enforce selected configuration expectations, such as RBAC usage, network exposure or protection features, using built-in or custom definitions where supported. Policy evaluation complements authorization: it describes expected resource state, while RBAC defines what an identity may do. Their responsibilities should not be collapsed into one vague security requirement.
Choose effects carefully. Audit can provide visibility before enforcing a change; Deny can stop noncompliant deployments, but it can also disrupt necessary remediation if the rule or exemption is poorly scoped. A staged rollout should inventory vaults, classify exceptions, identify service owners and demonstrate that approved deployments still work. The existing Azure Policy initiative coverage is useful when organizing controls into a repeatable subscription-level standard.
Exceptions are sometimes unavoidable for migrations or specialized networking arrangements, but they need a business rationale, owner, expiration and compensating controls. Do not accept a permanent broad firewall exception just because the application team cannot yet explain the DNS route. Recheck deviations after architecture changes, since a temporary exception can become a persistent route to sensitive material.
Connect Defender for Key Vault alerts to investigation and verification
Microsoft Defender for Key Vault offers threat-detection signals for suspicious or anomalous access patterns. It is a different layer from RBAC, a firewall or a secret rotation policy. Those preventive controls decide or limit allowed use; threat detection helps determine whether activity warrants investigation after it occurs. Alert coverage depends on the relevant plan and configuration, so verify that the expected vaults are included before using a dashboard as evidence of protection.
An investigation should establish the identity or network source, vault and object scope, requested action, timing, known maintenance activity and any neighboring identity compromise. A legitimate automation job can resemble suspicious bulk access; conversely, a familiar principal may have been hijacked. The Defender for Key Vault documentation emphasizes investigating alert context rather than treating recognition of an account as sufficient reassurance.
Log access and ownership need their own operational controls. Keep necessary diagnostic data available to the responsible team and protect the log destination. Test whether alerts reach the on-call workflow, whether responders can identify the affected application, and whether the chosen remediation actually closes the unsafe path. A chart showing that an alert was generated is not the same as evidence that it was triaged and resolved.
Verify the entire access journey before signing off
The final test should walk from a real workload identity through token acquisition, role authorization, network and DNS resolution, object retrieval, monitoring and failure response. Confirm both positive and negative cases: authorized retrieval succeeds; an unapproved identity is denied; a disallowed network is blocked; deleted or rotated material follows the expected recovery path. Record expected error conditions so a later change can be evaluated against them.
Use a scenario in which a production application is moved to a different subnet while its Key Vault remains private. Test whether the new route resolves correctly, whether the application still uses the right identity, and whether an operations engineer can diagnose a failure without temporarily granting broad credentials. Then simulate a suspicious vault access signal and check incident ownership. This exposes the boundaries that a portal-only checklist misses.
The Cloud and AI Security Engineer Associate role rewards engineers who can turn those tests into stable controls. Key Vault is secure when permission, connectivity, lifecycle and detection make sense together—not merely when a secret appears inside a managed service.