RMISC 2024 - Exploitable Weaknesses in Gen AI Workflows: From RAG to Riches

This presentation was delivered at the Rocky Mountain Information Security Conference (RMISC) in June 2024, but the talk was not recorded. Attendees asked if they could find the material elsewhere and so we're posting this recording from a few days after the conference.
Original abstract:
Everyone's building AI chatbots using Retrieval Augmented Generation (RAG) with Large Language Models (LLM), but how many of these teams understand the risks they're opening themselves up to, especially as they mix confidential data with new types of databases and other infrastructure. This session will demonstrate attacks on the "memory of AI," vector databases, which are used in countless ways from RAG to facial recognition to medical diagnoses. The AI data is a treasure trove for attackers. We'll end by showing how to defend against these completely new attacks.