Quick answer: Azure Key Vault can supply secrets to AI workloads using managed identities or other authenticated principals and Azure access controls. Separate identities, secret permissions and runtime isolation matter more than whether a workload is called an agent. Gazebo offers a separate stored-credential retrieval workflow, not a live Key Vault source connector.
What Key Vault controls
Key Vault protects access to keys, secrets and certificates. These are different resource types: an HSM-backed cryptographic key is not the same thing as an arbitrary SaaS secret stored in the vault.
Microsoft Entra identities authenticate workloads. Azure RBAC can authorize data-plane access, while existing deployments may use the legacy access-policy model. Microsoft's authorization guidance recommends RBAC; review the actual model and inherited grants in your vault.
Distinguish metadata visibility from secret values. A role such as Key Vault Reader is not interchangeable with Key Vault Secrets User. Also distinguish permission to administer the vault from permission to consume a secret.
An agent that retrieves a GitHub token through Key Vault then uses that token at GitHub. Key Vault controls the retrieval; GitHub controls the token's repository and operation permissions.
Identity does not have to mean one shared host
A common deployment gives several processes access to the same managed identity. Those processes may then have equivalent Key Vault access. This is a shared-identity boundary, not proof that Azure cannot distinguish agent workloads.
Separate deployments can use distinct identities. Appropriate managed or federated workload identities and scoped role assignments can provide different secret access for different workloads. Review the permissions and isolation provided by the hosting platform rather than assuming a new profile name separates processes.
Assigning several identities to a resource also does not automatically prevent code on that resource from using another identity. The runtime's actual access to token endpoints, files, tools and neighbouring processes determines the effective boundary.
For an explanation of identity versus permissions, see IAM for AI agents.
A native Azure workflow
For an Azure-hosted agent that needs a sandbox provider key:
- Identify the runtime. Determine which principal the actual SDK/runtime will use, not merely the identity configured in a design document.
- Restrict secret retrieval. Grant the required data-plane permissions at an appropriate scope. Review inherited role assignments and use Microsoft's recommended vault/resource organization.
- Restrict the provider key. Choose a sandbox or provider-restricted credential. A read-only Key Vault grant does not make a returned key read-only at its provider.
- Configure diagnostics. Send the needed Key Vault audit events to protected destinations with suitable retention and access controls.
- Test the boundaries. Confirm the intended read succeeds, an unrelated secret is denied and provider-side permissions block prohibited actions.
- Test lifecycle changes. Exercise identity removal, credential replacement, application reload and retained-key containment separately.
For secret-level role assignments, check Microsoft's guidance on appropriate use rather than making thousands of individual assignments the default architecture.
Retrieval evidence is not action evidence
Key Vault logging requires diagnostic configuration and a destination such as storage or Log Analytics. Check that events actually arrive, that the identity context is useful and that operators can access them during an incident.
A secret-access record is evidence about the vault request. It does not show every Stripe refund, GitHub change or database query performed with the returned credential. Provider and runtime evidence are separate and may have their own availability, retention and completeness limits.
Likewise, an agent saying “I only read customer data” is not equivalent to a trusted execution record. Do not log raw credentials or unrestricted request/response bodies in an attempt to fill the gap.
Rotation and removal
Managed identity reduces the need to distribute bootstrap secrets for supported authentication paths. It does not automatically rotate every third-party credential kept as a Key Vault secret.
Separate key-rotation capabilities, certificate renewal and external-secret rotation workflows. Confirm the mechanism for the particular resource, who updates the provider, and how consumers adopt the replacement.
Removing an identity's secret-read permission blocks future reads through that authorization path, subject to the platform's authorization behaviour. It does not invalidate a provider token already copied into a runtime. Containing that copy may require restriction, revocation or expiry at its provider.
If several workloads share one provider key, rotating it can affect all of them even when their Key Vault identities are distinct. Plan replacement and verification accordingly.
Where Gazebo fits today
Gazebo's potential advantage is its focused workflow for connecting agent clients, approving credential retrieval and reviewing access across services. Its integration tooling can help custom clients use that path without inventing a new secret-delivery protocol.
Gazebo currently stores the credentials it releases in its own vault. It does not fetch live values from Azure Key Vault or preserve Azure RBAC automatically. Using both stores requires an explicit provisioning and update process, with a named owner for each copy.
Gazebo checks a requesting agent's service and declared-method retrieval policy; OAuth connections also approve specific service keys. An authorized release delivers the provider credential to the runtime and host. It does not enforce later endpoints, records, payment operations or task intent.
Use provider permissions and a trusted executing system for those controls. Gazebo retrieval history can complement—not replace—provider activity evidence.
Decide based on the actual workflow
Azure-native access may be sufficient for agents deployed alongside existing services. Gazebo may be useful when several independent agent clients need a central, understandable retrieval experience across providers.
Before adding it, compare the setup benefit against another credential copy, another administrative boundary and the update process required. Test removal of alternate access paths too: routing one tool through Gazebo does not remove a key still available through Key Vault, a shell environment or another tool.
See agent credential management for the lifecycle and least privilege for effective-access testing.
Independent editorial post; not affiliated with or endorsed by Microsoft. Last reviewed: 8 November 2026. Consult current Azure documentation for identity, authorization and feature-specific configuration.