The era of simply “slapping an LLM on a database” is over. Retrieval-Augmented Generation (RAG) solved AI hallucinations, but it introduced a massive, highly complex attack surface. When you give an LLM access to internal company documents, vector databases, and API tools (Agentic RAG), you are effectively turning passive data into executable code.
Without proper defenses, a single poisoned PDF or a hidden prompt injection can lead to data exfiltration, Server-Side Request Forgery (SSRF), or a complete system compromise.
In this dense, zero-fluff, 90-minute masterclass, AI Security Researcher Armaan Sidana takes you deep into the trenches of offensive AI security. You won’t just learn high-level theory—you will actively hack Vector Databases, execute context hijacking, and manipulate AI agents. Then, you will learn how to build the ultimate 4-Gate Defense Architecture to lock down your pipelines for production.
What You Will Learn:
- Vector Database Exploitation: Discover why default Vector DBs are the soft underbelly of AI. Execute a live heist on a vulnerable Qdrant instance and understand the dangers of Embedding Inversion (turning math back into raw PII).
- Advanced Data Poisoning: Hack the ingestion pipeline. Learn how attackers use invisible text (Font-Size 0), Unicode steganography, and metadata injection to poison RAG systems.
- Context Hijacking & Overflow: Exploit the LLM’s “U-Shaped Attention” mechanism and execute Context Stuffing attacks that push safety instructions entirely out of the memory window.
- Agentic RAG & SSRF: Watch what happens when RAG pipelines grow hands. Trick AI agents into acting as a “Confused Deputy” to leak cloud credentials or abuse internal APIs.
- The 4-Gate Defense Architecture: Build a bulletproof system. Implement Magic Byte validation, Semantic Chunking, Meta’s Prompt Guard, XML Sandboxing, and strict Grounding Evaluators (LLM-as-a-Judge).
- Automated Red Teaming: Stop doing manual penetration testing. Learn how to deploy Promptfoo in your CI/CD pipelines, use NVIDIA Garak for deep fuzzing, and orchestrate stateful attacks with Microsoft PyRIT.
Real-World Case Studies & CTF
Learn from the costly mistakes of others. We will deconstruct major real-world failures, including the Air Canada hallucinated policy lawsuit, the Bing “Sydney” prompt leak, and the GitHub Copilot RCE vulnerability.
Finally, put your skills to the test in the Capstone CTF (Capture The Flag), where you will use cross-language translation and context manipulation to break out of a restricted RAG agent and steal an admin password.
