Adversarial Defenses for LLMs

​In his talk, Samuel Simko from ETH Zurich will present his recent work on adversarial defenses for LLMs, developed with the Jinesis Lab (University of Toronto). The talk will cover a series of approaches, ranging from triplet-based contrastive learning defenses to honeypot-style defenses designed to avoid worst-case behavior. He will also discuss patterns observed in contest-winning manual jailbreaking prompts, ideas for tamper-resistant safeguards, and the current limits of attacks, defenses, and evaluation methodologies.