PaperGPT : Jailbreaking Black Box LLMs-Jailbreak LLMs Efficiently

Introduction to PaperGPT : Jailbreaking Black Box LLMs

PaperGPT : Jailbreaking Black Box LLMs, also known as 'PAIR' (Prompt Automatic Iterative Refinement), is an algorithm designed to generate semantic jailbreaks with only black-box access to a large language model (LLM). Inspired by social engineering tactics, PAIR employs an attacker LLM to automatically produce jailbreaks for a targeted LLM. This iterative process requires fewer than twenty queries to potentially elicit a jailbreak, significantly enhancing efficiency over previous methods. An example of PAIR in action involves it automatically generating prompts that, when fed into a target LLM, produce responses that breach preset ethical or safety guidelines, typically within a constrained number of queries. Powered by ChatGPT-4o。