Retrieval-Augmented Generation (RAG) is becoming the backbone of enterprise AI and yet it also creates one of today’s most exploited attack surfaces. Beyond jailbreaks, attackers now poison documents, corrupt embeddings, hijack retrieval, and manipulate prompt context to silently steer outputs. This session reveals the full, threat model behind RAG attacks, from ingestion poisoning and vector drift to cross-document injections, Unicode traps, semantic misdirection, and multi-agent identity creep.

I will show you how to secure Azure AI Search ingestion, Fabric governance, Entra ID identity boundaries, Defender for Cloud Apps telemetry, Sentinel anomaly detection, and Copilot Studio isolation patterns.
Using a live-style breakdown, I will contrast how a compromised RAG behaves under attack (and how the same pipeline becomes resilient when properly secured). You'll leave knowing how to design RAG systems that not only work, but survive real adversaries.