IC-929Simulated character-level ciphers that never appear in pretraining data cannot bypass safety alignment even with 10+ demonstrations

Youliang Yuan, Wenxiang Jiao, Wenxuan Wang, Jen-tse Huang, Pinjia He, Shuming Shi, Zhaopeng Tu

SourceGPT-4 Is Too Smart To Be Safe: Stealthy Chat with LLMs via Cipher

The authors define non-existent ciphers using random alphabet mapping and Chinese character substitutions that have no prior in the models' pretraining data. These simulated ciphers fail to produce valid or unsafe responses even when provided with more than 10 in-context demonstrations. This contrasts with human ciphers (Caesar, ASCII) and SelfCipher, which do work, indicating that the models' cipher comprehension relies on character-level priors learned during pretraining rather than on-the-fly rule learning.

Evidence
correlational
Caveat
The paper does not report specific numbers for the simulated cipher experiments; the result is stated qualitatively. The authors note that self-defined word-level ciphers (from a separate study) can work, suggesting the limitation is specific to character-level ciphers.
Model
GPT-4 / ChatGPT4 / GPT-4 Code Interpreter / GPT-4 Technical Report, GPT-3.5 / ChatGPT-3.5 GPT-3.5-turbo
Datasets
Safety Assessment of Chinese Large Language Models [eval]
Related work
Jailbreaking proprietary large language models using word substitution cipher [compared-to]
Related findings
IC-927, IC-928
Extraction
automatic-extraction