Source: Universal and Transferable Adversarial Attacks on Aligned Language Models (GCG)
Zou et al. (2023): greedy plus gradient search finds adversarial suffixes on open models that transfer to ChatGPT, Bard and Claude.
Contents
- What: research paper, arXiv:2307.15043 (method commonly called Greedy Coordinate Gradient, GCG)
- Authors: Andy Zou, Zifan Wang, Nicholas Carlini, Milad Nasr, J. Zico Kolter, Matt Fredrikson
- Date: 2023-07-27 (v2 2023-12-20)
- Original: arXiv:2307.15043
Summary
Automates jailbreaks by optimising a token suffix appended to a harmful request so that the model’s most likely continuation begins affirmatively (“Sure, here is…”) rather than refusing. Search combines greedy and gradient-based token substitution, requiring white-box access to open models.
Key claims
- Suffixes optimised over many prompts and over Vicuna-7B and 13B transfer to black-box systems: ChatGPT, Bard and Claude public interfaces, and to LLaMA-2-Chat, Pythia and Falcon.
- Suffixes are automatically generated, not hand-written, and often look like gibberish.
- Code: github.com/llm-attacks/llm-attacks.