Safety Assessment of Chinese Large Language Models

Report a problem with this record

Findings

IC-927Human ciphers (ASCII, Unicode, Caesar, Morse) bypass the safety alignment of GPT-4 and GPT-3.5-turbo, with more powerful models producing more unsafe responses [eval]
IC-928SelfCipher (a role-play prompt without explicit cipher rules) evokes a 'secret cipher' in LLMs, achieving high unsafety rates that outperform most human ciphers [eval]
IC-929Simulated character-level ciphers that never appear in pretraining data cannot bypass safety alignment even with 10+ demonstrations [eval]