Jailbreaking LLMs Through Cross-Cultural Prompts

Damin Kim; Minseok Hur; Jeongin Lee; Moohong Min

doi:10.1145/3746252.3760892

Jailbreaking LLMs Through Cross-Cultural Prompts

Damin Kim
, Minseok Hur
, Jeongin Lee
, Moohong Min

University College

Sungkyunkwan University

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › peer-review

Abstract

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.

Original language	English
Title of host publication	CIKM 2025 - Proceedings of the 34th ACM International Conference on Information and Knowledge Management
Publisher	Association for Computing Machinery, Inc
Pages	4874-4878
Number of pages	5
ISBN (Electronic)	9798400720406
DOIs	https://doi.org/10.1145/3746252.3760892
State	Published - 10 Nov 2025
Event	34th ACM International Conference on Information and Knowledge Management, CIKM 2025 - Seoul, Korea, Republic of Duration: 10 Nov 2025 → 14 Nov 2025

Publication series

Name	CIKM 2025 - Proceedings of the 34th ACM International Conference on Information and Knowledge Management

Conference

Conference	34th ACM International Conference on Information and Knowledge Management, CIKM 2025
Country/Territory	Korea, Republic of
City	Seoul
Period	10/11/25 → 14/11/25

Keywords

alignment
cross-cultural prompts
jailbreaking
large language models

Access to Document

10.1145/3746252.3760892

Cite this

Kim, D., Hur, M., Lee, J., & Min, M. (2025). Jailbreaking LLMs Through Cross-Cultural Prompts. In CIKM 2025 - Proceedings of the 34th ACM International Conference on Information and Knowledge Management (pp. 4874-4878). (CIKM 2025 - Proceedings of the 34th ACM International Conference on Information and Knowledge Management). Association for Computing Machinery, Inc. https://doi.org/10.1145/3746252.3760892

Kim, Damin ; Hur, Minseok ; Lee, Jeongin et al. / Jailbreaking LLMs Through Cross-Cultural Prompts. CIKM 2025 - Proceedings of the 34th ACM International Conference on Information and Knowledge Management. Association for Computing Machinery, Inc, 2025. pp. 4874-4878 (CIKM 2025 - Proceedings of the 34th ACM International Conference on Information and Knowledge Management).

@inproceedings{b54612409cac483a801b9aa3c6e8cf17,

title = "Jailbreaking LLMs Through Cross-Cultural Prompts",

abstract = "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.",

keywords = "alignment, cross-cultural prompts, jailbreaking, large language models",

author = "Damin Kim and Minseok Hur and Jeongin Lee and Moohong Min",

note = "Publisher Copyright: {\textcopyright} 2025 Copyright held by the owner/author(s).; 34th ACM International Conference on Information and Knowledge Management, CIKM 2025 ; Conference date: 10-11-2025 Through 14-11-2025",

year = "2025",

month = nov,

day = "10",

doi = "10.1145/3746252.3760892",

language = "English",

series = "CIKM 2025 - Proceedings of the 34th ACM International Conference on Information and Knowledge Management",

publisher = "Association for Computing Machinery, Inc",

pages = "4874--4878",

booktitle = "CIKM 2025 - Proceedings of the 34th ACM International Conference on Information and Knowledge Management",

}

Kim, D, Hur, M, Lee, J & Min, M 2025, Jailbreaking LLMs Through Cross-Cultural Prompts. in CIKM 2025 - Proceedings of the 34th ACM International Conference on Information and Knowledge Management. CIKM 2025 - Proceedings of the 34th ACM International Conference on Information and Knowledge Management, Association for Computing Machinery, Inc, pp. 4874-4878, 34th ACM International Conference on Information and Knowledge Management, CIKM 2025, Seoul, Korea, Republic of, 10/11/25. https://doi.org/10.1145/3746252.3760892

Jailbreaking LLMs Through Cross-Cultural Prompts. / Kim, Damin; Hur, Minseok; Lee, Jeongin et al.
CIKM 2025 - Proceedings of the 34th ACM International Conference on Information and Knowledge Management. Association for Computing Machinery, Inc, 2025. p. 4874-4878 (CIKM 2025 - Proceedings of the 34th ACM International Conference on Information and Knowledge Management).

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › peer-review

TY - GEN

T1 - Jailbreaking LLMs Through Cross-Cultural Prompts

AU - Kim, Damin

AU - Hur, Minseok

AU - Lee, Jeongin

AU - Min, Moohong

PY - 2025/11/10

Y1 - 2025/11/10

N2 - 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.

AB - 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.

KW - alignment

KW - cross-cultural prompts

KW - jailbreaking

KW - large language models

UR - https://www.scopus.com/pages/publications/105023147674

U2 - 10.1145/3746252.3760892

DO - 10.1145/3746252.3760892

M3 - Conference contribution

AN - SCOPUS:105023147674

T3 - CIKM 2025 - Proceedings of the 34th ACM International Conference on Information and Knowledge Management

SP - 4874

EP - 4878

BT - CIKM 2025 - Proceedings of the 34th ACM International Conference on Information and Knowledge Management

PB - Association for Computing Machinery, Inc

T2 - 34th ACM International Conference on Information and Knowledge Management, CIKM 2025

Y2 - 10 November 2025 through 14 November 2025

ER -

Kim D, Hur M, Lee J, Min M. Jailbreaking LLMs Through Cross-Cultural Prompts. In CIKM 2025 - Proceedings of the 34th ACM International Conference on Information and Knowledge Management. Association for Computing Machinery, Inc. 2025. p. 4874-4878. (CIKM 2025 - Proceedings of the 34th ACM International Conference on Information and Knowledge Management). doi: 10.1145/3746252.3760892

Jailbreaking LLMs Through Cross-Cultural Prompts

Abstract

Publication series

Conference

Keywords

Access to Document

Other files and links

Fingerprint

Cite this