Robust Generalization in the Era of LLMs - Jailbreaking Attacks and Defenses
Explore the current landscape of jailbreaking attacks and defenses in large language models (LLMs) through this informative lecture by Hamed Hassani from the University of Pennsylvania. Delve into the vulnerabilities of popular LLMs like GPT, Llama, Claude, and Gemini to adversarial manipulation, and examine the growing interest in improving their robustness. Gain insights into the latest developments in the jailbreaking literature, including new perspectives on robust generalization, innovative black-box attacks on LLMs, and emerging defense strategies. Learn about a new leaderboard designed to evaluate the robust generalization capabilities of production LLMs. Understand the challenges and opportunities in aligning LLMs with human intentions and protecting them against malicious exploitation.