Entra Workload Identity Reference

Reference for Microsoft Entra Workload Identity. The API variables are listed with the other API environment variables, and the AI system fields are described in Custom API Language Models.

Availability

Capability	Platform Release
Entra workload identity for Custom API Language Models	3.26.8
Entra for judge, attacker, and evaluation data generation calls	3.26.8
Entra for DynamoGuard policy data generation	3.26.9
LLM_EXTRA_HEADERS for DynamoGuard policy data generation	3.26.10
A customer-owned identity per AI system	3.26.11

Identity Used by Each Call

Call	Token Is Minted As	Configured On
Requests to the AI system under evaluation	The platform identity, or the customer-owned identity named on the AI system	The AI system
Connection test when the AI system is saved	The platform identity; skipped when the AI system names its own identity	The AI system
Judge and attacker models, and evaluation data generation	The platform identity	Evaluation worker variables
DynamoGuard policy data generation	The platform identity	Data processing variables

Tokens are requested for the configured scope and refreshed before they expire. The APIM subscription key, when set, is sent as ocp-apim-subscription-key on every call.

Save-Time Behavior

What happens when an AI system that uses Microsoft Entra Workload Identity is created, or its authentication is edited, and how it then behaves during evaluation.

Managed Identity Client ID and Tenant ID	On Save	During Evaluation
Neither set (a blank value counts as not set)	The API requests a token as the platform identity, sends a test request, and checks the response against the response transform	Tokens are minted as the platform identity
Both set	The configuration, the endpoint host, and the request transform are checked. No token is requested and no test request is sent	Tokens are minted as the named identity; an authentication problem fails the first run
Only one set	Rejected: config.client_id and config.tenant_id must be provided together	Not reached
A value that is not a GUID	Rejected: config.<field> must be a GUID if provided	Not reached
Any other configuration key, such as token_file_path	Rejected: config contains invalid fields: <field>	Not reached
Edit that removes both	The AI system returns to the platform identity, and the test request runs again	Tokens are minted as the platform identity
Edit that leaves the APIM subscription key empty	The stored key is kept	The stored key is sent

The authentication type of an existing AI system cannot be changed. Create a new AI system to move one from Bearer Token or API Key to Microsoft Entra Workload Identity.

Evaluation Worker Variables

Set these on each evaluation worker job in the DynamoEval chart to route judge, attacker, and evaluation data generation calls to an Entra-protected Azure endpoint. The workers read AZURE_CLIENT_ID, AZURE_TENANT_ID, and AZURE_FEDERATED_TOKEN_FILE from the workload identity webhook.

Variable	Required	Description
DYNAMOEVAL_JUDGE_MODE	Yes	AZURE routes judge and attacker calls to the endpoint below.
AZURE_API_BASE	Yes	Endpoint base URL, for example the APIM URL in front of Azure OpenAI.
AZURE_API_VERSION	Yes	Azure OpenAI API version.
AZURE_MODEL_NAME	No	Bare deployment name. Routes every judge, attacker, and evaluation data generation call to this deployment.
AZURE_ENTRA_AUTH	Yes, for Entra	true mints Entra tokens instead of using AZURE_API_KEY. The two are mutually exclusive; leave AZURE_API_KEY unset.
AZURE_ENTRA_CREDENTIAL	No	workload_identity (default), or client_secret, which also reads AZURE_CLIENT_SECRET.
AZURE_ENTRA_SCOPE	No	Scope the token is requested for. Defaults to https://cognitiveservices.azure.com/.default; it must match the audience the gateway validates.
AZURE_APIM_SUBSCRIPTION_KEY	No	APIM subscription key sent with every call. Read it from the azureApimSubscriptionKey key of the common Secret.
AZURE_APIM_SUBSCRIPTION_KEY_HEADER	No	Header name for the key. Defaults to ocp-apim-subscription-key.

Entra for judge and attacker calls on an evaluation worker job

dynamoai-common:
  scaledJobs:
    jobs:
      <job-name>:
        env:
          DYNAMOEVAL_JUDGE_MODE:
            value: "AZURE"
          AZURE_API_BASE:
            value: "https://<gateway-host>/<api-path>"
          AZURE_API_VERSION:
            value: "<api-version>"
          AZURE_MODEL_NAME:
            value: "<deployment>"
          AZURE_ENTRA_AUTH:
            value: "true"
          AZURE_APIM_SUBSCRIPTION_KEY:
            valueFrom:
              secretKeyRef:
                name: "{{ .Values.global.secrets.common }}"
                key: azureApimSubscriptionKey

How evaluation data generation picks its models is covered in DynamoEval Workers.

Data Processing Variables

Set these on the data processing service to generate DynamoGuard policy data through an Entra-protected Azure OpenAI endpoint.

Variable	Required	Description
SYNDATA_MODEL_PROVIDER	Yes	azure_openai.
SYNDATA_MODEL_NAME	Yes	Provider-prefixed deployment, for example azure/<deployment>. In the DynamoGuard chart it normally takes the value of global.models.dataGeneration.name.
AZURE_OPENAI_ENDPOINT	Yes	Azure OpenAI endpoint URL.
AZURE_OPENAI_API_VERSION	Yes	Azure OpenAI API version.
AZURE_ENTRA_AUTH	Yes, for Entra	true mints Entra tokens instead of using AZURE_OPENAI_API_KEY. The two are mutually exclusive; leave AZURE_OPENAI_API_KEY unset.
AZURE_ENTRA_CREDENTIAL, AZURE_ENTRA_SCOPE, AZURE_APIM_SUBSCRIPTION_KEY, AZURE_APIM_SUBSCRIPTION_KEY_HEADER	No	As for the evaluation workers.
LLM_EXTRA_HEADERS	No	JSON object of extra headers sent on every LLM call, for example {"x-end-user": "dynamoai"}.