Question

A financial services company is adopting Azure OpenAI to build an internal customer-support assistant that will access sensitive customer records. The CISO wants the architecture to follow Microsoft's guidance for secure AI adoption within the Cloud Adoption Framework. Leadership is specifically concerned about preventing the model from returning data a user is not authorized to see, and about detecting prompt-injection and data-exfiltration attempts. Which design approach best aligns with Microsoft's secure AI adoption guidance for these concerns?

Accepted Answer

Enforce identity-based authorization at the data retrieval layer (grounding data) using the calling user's context, and deploy Microsoft Defender for Cloud AI threat protection plus Azure AI Content Safety prompt shields to detect prompt-injection and exfiltration attempts.

Answer

Rely on the Azure OpenAI content filters alone, since they block harmful content and therefore also enforce data authorization boundaries.

Answer

Grant the application's managed identity broad read access to all customer records and instruct the system prompt to only reveal data relevant to the authenticated user.

Answer

Fine-tune the model on the full customer dataset so that authorization logic is embedded in the model weights, removing the need for runtime access controls.