how do i secure my Azure AI Foundry deployed models.

Hello @Sen0299 ,

Welcome to Microsoft Q&A .Thank you for reaching out to us.

Securing Azure AI Foundry deployed models accessed through APIs is best handled through a layered approach. The most effective path is to begin with identity and access controls, then add network protection, monitoring, content safety and API governance based on workload sensitivity and production requirements.

Following are the suggested approaches-

1. Strengthening authentication and access control The first layer is to control which identities can access the model endpoint. Recommended actions:
   1. Use Entra ID authentication and Managed Identity wherever supported, instead of relying only on API keys.
   2. Apply least-privilege RBAC, granting only the permissions required for inference or resource access.
   3. Avoid broad roles such as Owner or Contributor unless administration access is required.
   4. If API keys are still required:
      • Store them securely in Azure Key Vault.
      • Rotate keys regularly.
      • Avoid placing keys in source code, client-side applications, configuration files, or logs.
      This reduces the risk of leaked credentials, but it is still important to note that RBAC and identity controls restrict access; they do not fully prevent misuse by an already-authorized caller.
2. Restricting network exposure After access control, the next layer is limiting where API calls can originate. Recommended actions:
   1. Replace All networks access with Selected Networks/IP allow lists where possible.
   2. For production or sensitive workloads, configure Private Endpoint / Private Link.
   3. Disable public network access when private access is required.
   4. Validate DNS and connectivity after Private Endpoint configuration.
   Selected Networks helps reduce exposure, but the public endpoint may still exist. Private Endpoint provides stronger isolation, although it adds networking, DNS, and cost considerations.
3. Adding content safety and abuse protection For applications accepting user-generated prompts, content and abuse controls should be included. Recommended actions:
   1. Use the default content filtering available for Azure OpenAI deployments.
   2. Add Azure AI Content Safety for stronger input/output protection.
   3. Consider Prompt Shields, jailbreak detection, and custom blocklists for public-facing or high-risk applications.
   These controls help reduce harmful content, prompt injection, and abuse risks.
4. Using API governance for production scenarios For multiple applications, teams, or consumers, placing Azure API Management AI Gateway in front of model endpoints is recommended. This provides:
   • Centralized authentication
   • Rate limiting
   • Token quotas
   • Traffic throttling
   • Policy enforcement
   • Usage monitoring
   • Cost control
   • Optional prompt and completion logging
   This is especially useful for preventing one application or user from consuming excessive quota or causing unexpected cost increases.

As for included vs additional-cost options

The following references might be helpful , please check them out

1. Included / foundational controls
   1. Entra ID authentication
   2. Managed Identity
   3. RBAC
   4. Selected Networks/IP restrictions
   5. Azure Monitor metrics
   6. Default content filtering for Azure OpenAI deployments
2. Additional-cost / advanced controls
   1. Private Endpoint / Private Link
   2. Azure API Management AI Gateway
   3. Azure AI Content Safety advanced usage
   4. Log Analytics ingestion
   5. Defender for Cloud AI threat protection
   6. Azure Key Vault usage and transactions

For token usage and prompt/response visibility

Azure Monitor can show operational and token metrics such as:

• Processed Prompt Tokens
• Generated Completion Tokens
• Processed Inference Tokens
• Request volume
• Latency
• Availability

These metrics help with usage tracking and consumption analysis.

However, Azure AI Foundry deployment metrics do not automatically provide a built-in portal view of every historical prompt and completion sent through a resource API key. Metrics show usage and performance data, not full request and response bodies by default

If actual prompt and response text must be captured for auditing or troubleshooting, practical options include:

1. Implement application-level logging before and after model calls, with sensitive data redaction.
2. Route calls through Azure API Management AI Gateway and enable LLM logging to capture prompts, completions, model details, and token usage.
3. For supported agent scenarios, enable Foundry Tracing to capture prompt content, conversation history, inputs, outputs, token details, and tool activity Detailed prompt and response logging may contain sensitive information, so access control, retention, redaction, and compliance review should be applied before enabling it

As a data privacy note,for models sold by Azure in Azure AI Foundry, prompts, completions, embeddings, and training data are not available to other customers, are not available to OpenAI or other model providers, and are not used to train foundation models without permission or instruction.

These are the suggested Implementation steps in order

1. Configure Entra ID / Managed Identity.
2. Apply least-privilege RBAC.
3. Restrict access using Selected Networks/IP allow lists.
4. Enable Azure Monitor metrics.
5. Add Content Safety controls where applicable.
6. Configure Private Endpoint and disable public access for sensitive workloads.
7. Place Azure API Management AI Gateway in front of the model endpoint.
8. Add Defender for Cloud and advanced monitoring where required.

With these controls in place, the model deployment can be secured at the identity, network, monitoring, governance and data-protection layers while still allowing the required API-based access.

The following references might be helpful , please check them out

• Azure AI security best practices | Microsoft Learn
• Azure security baseline for Microsoft Foundry | Microsoft Learn
• How to configure network isolation for Microsoft Foundry - Microsoft Foundry | Microsoft Learn
• AI gateway capabilities in Azure API Management | Microsoft Learn
• Set Up Logging for Language Model APIs - Azure API Management | Microsoft Learn
• Monitoring data reference for Azure OpenAI - Microsoft Foundry | Microsoft Learn
• Set Up Tracing for AI Agents in Microsoft Foundry - Microsoft Foundry | Microsoft Learn
• Data, privacy, and security for Foundry Models sold by Azure in Microsoft Foundry - Microsoft Foundry | Microsoft Learn

Please let us know if the response was helpful

Thank you