Exploiting Shadow Data from AI Models and Embeddings

Learn how to exploit hidden vulnerabilities in AI-powered applications through this DEF CON 33 conference talk that exposes the privacy risks lurking in modern AI systems. Discover how sensitive data including personally identifiable information (PII) and social security numbers can be extracted from large language models (LLMs) and retrieval-augmented generation (RAG) workflows through real-world attack demonstrations. Explore model inversion attacks that target fine-tuned models and embedding inversion attacks on vector databases, while understanding why traditional PII scanning tools fail to detect the rich data stored within these AI ecosystems. Gain insights into the significant privacy disasters that AI systems can create and the shadow data vulnerabilities that organizations often overlook when implementing AI-powered applications.