Source: Jailbreaking Black Box Large Language Models in Twenty Queries (PAIR)
Chao et al. (2023): an attacker LLM iteratively refines semantic jailbreak prompts against a black-box target, often in under 20 queries.
Contents
- What: research paper, arXiv:2310.08419
- Authors: Patrick Chao, Alexander Robey, Edgar Dobriban, Hamed Hassani, George J. Pappas, Eric Wong
- Date: 2023-10-12 (v4 2024-07-18)
- Original: arXiv:2310.08419
Summary
PAIR (Prompt Automatic Iterative Refinement) uses one LLM as an attacker that proposes a jailbreak, reads the target’s response, and refines the prompt, inspired by social engineering. It needs only black-box access and produces human-readable prompts.
Key claims
- Often finds a jailbreak in fewer than twenty queries, orders of magnitude more efficient than token-level search.
- Competitive success and transferability on GPT-3.5/4, Vicuna and Gemini.