Reproduce “Obscure but Effective: Classical Chinese Jailbreak Prompt Optimization via Bio-Inspired Search” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xun Huang, Simeng Qin, Xiaoshuang Jia, Ranjie Duan, Huanqian Yan, Zhitao Zeng, Fei Yang, Yang Liu and 1 more
Researchers introduced CC-BOS, a black-box jailbreak framework that generates adversarial prompts in Classical Chinese, to expose a novel linguistic vulnerability in state-of-the-art Large Language Models. Leveraging a multi-dimensional strategy space and a bio-inspired optimization algorithm, the framework achieved a 100% attack success rate on the AdvBench benchmark across six mainstream LLMs, significantly outperforming existing jailbreak methods in both effectiveness and efficiency.
Key insights
- Classical Chinese's semantic compression, ambiguity, and rich rhetorical techniques create a 'safety blind spot' in LLM defenses, allowing malicious content to bypass safeguards.
- A systematic, multi-dimensional strategy space is effective for constructing complex, stealthy adversarial prompts that leverage specific linguistic and cultural nuances.
- Bio-inspired optimization algorithms can efficiently discover highly effective jailbreak prompts in a black-box setting across intricate linguistic contexts, indicating a 'High Capability-Low Alignment' vulnerability in LLMs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.