On the Hardness of Junking LLMs

Marco Rando, Samuel Vaiter

Large language models (LLMs) are known to be vulnerable to jailbreak attacks, which typically rely on carefully designed prompts containing explicit semantic structure. These attac…