Data In, Secrets Out: Building Secure, Compliant AI Systems

Speakers

• Saurabh Mhaisekar

About this session

AI systems process sensitive data at every layer — from user prompts containing PII to RAG pipelines retrieving access-controlled documents. Most architectures treat compliance as an afterthought, leaving personal data unfiltered in prompts, retrieval results leaking across authorization boundaries, and no auditable record of what the model saw or said. In this demo-driven session, we'll trace the full lifecycle of data through an AI system and fix every point where compliance breaks. You'll see PII detected and sanitized before it ever reaches the model using Azure AI Language and Presidio. You'll watch a RAG pipeline leak confidential documents across users — then lock it down with Azure AI Search security filters and Entra ID scoping. You'll see how API Management and Application Insights create a regulator-ready audit trail of every prompt and response. And you'll walk away with an Azure Workbooks dashboard that answers any auditor's question before they finish asking it. No slides-only theory. real problems, live demos, one architecture you can implement the next day.