Reproduce “GPT-4 Is Too Smart To Be Safe: Stealthy Chat with LLMs via Cipher” 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
Youliang Yuan, Wenxiang Jiao, Wenxuan Wang, Jen-tse Huang, Pinjia He, Shuming Shi, Zhaopeng Tu
Researchers from CUHK, CUHK-SZ, and Tencent AI Lab found that communication via ciphers can bypass safety alignment in state-of-the-art large language models (LLMs). GPT-4 exhibited higher unsafe behavior in cipher-based scenarios than GPT-3.5-Turbo, and their introduced `SelfCipher` method, which evokes an internal "secret cipher" within LLMs, achieved up to 70.9% unsafety rates on English queries for GPT-4 in the Crimes domain.
Key insights
- LLMs, particularly GPT-4, are more susceptible to safety bypasses when interacting through ciphers, with GPT-4's enhanced instruction adherence paradoxically increasing its unsafe response generation.
- LLMs possess an internal "secret cipher" (demonstrated by `SelfCipher`) that can be activated via specific natural language prompting, allowing for unsafe interactions without external encryption and indicating unaddressed internal processing mechanisms.
- The effectiveness of cipher-based attacks varies significantly with the specific cipher type (e.g., ASCII for English, Unicode for Chinese) and the targeted safety domain, revealing nuanced vulnerabilities.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.