Breaking RAG Systems: Exploiting Vulnerabilities & Hardening Your GenAI Applicatio...
summary
A working RAG prototype uses OpenAI embeddings, GPT‑3.5‑turbo, Chroma vector store, LangChain QA chains and a Streamlit UI and is successfully broken via vector database poisoning (phishing, bad‑advice, redirect documents) and five prompt‑injection methods including role‑play and context stuffing and an encoding bypass; straight injection sometimes fails because model guardrails stop it. Fixes shown include digital signatures and multi‑approver ingestion, versioned audit trails, regex input sanitization, a validation layer and Falco anomaly detection, with code patterns that harden each failure mode.
AI-written summary · about
chapters (7)
- 0:00RAG vulnerabilities overview
- 3:41RAG system case study
- 8:20Basic RAG app demo
- 11:09Data/model poisoning attacks
- 16:24Prompt injection techniques
- 22:01Hardening defenses
- 24:29Q&A
related talks
41mFrom RAGs to riches: Build an AI document interrogation app in 30 mins
53mHow to Break AI Systems (Before Someone Else Does)
35mGenAI Application Security: Not Just Prompt Injection
51mFrom Text to Flaws: vulnerabilities in applications with Generative AI and LLMs
37mHow to Break AI Systems (Before Someone Else Does)
44mSecuring LLM-Powered Apps: Overcoming Security and Privacy Challenges
43mOne Poisoned Artifact Can Steer Your AI: How Robust Are Your LLM-Assisted Security Workflows?
40mSecuring Generative AI: Is it all an Illusion?