Stealing Production Language Models - Approaches and Risks

This conference talk from RSA Conference explores how adversaries can extract information from proprietary black-box AI systems like ChatGPT. Join UC Berkeley PhD Student Eric Wallace and Realm Labs CEO Saurabh Shintre as they examine two key attack vectors: exploiting LM output probabilities via API to reveal hidden model parameters, and training imitation models through distillation to create copycat systems. Discover the technical vulnerabilities of production language models and understand the implications for AI security in this 48-minute presentation from the cybersecurity experts at RSA Conference.