ScholarWorks@성균관대학교

초록

We examine how linguistic and cultural framing affect jailbreak success in three commercial LLMs (GPT-4, Claude 3, Gemini), using semantically equivalent prompts in direct, indirect, and metaphorical styles across four high-resource languages. Indirect prompts most effectively bypassed filters, with framing and style significantly influencing alignment. GPT-4 was especially vulnerable to indirect framing, Claude 3 remained consistently robust, and Gemini showed high sensitivity to cultural and linguistic variation. Our findings highlight the need for alignment strategies resilient to diverse expression styles and cultural contexts.

키워드