Gradient Cuff: Detecting Jailbreak Attacks on Large Language Models

Gradient Cuff proposes a two-step method to detect jailbreak attempts on Large Language Models by exploring refusal loss landscapes.