Analysis of Jailbreak Prompt Vulnerabilities in Commercial LLM
Journal of Data Forensics Research · 31.12.2024 · 1 (1) · 89-107
DOI: 10.12972/jdfr.2024.1.1.6 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of Data Forensics Research, online erschienen am 31.12.2024. Verfasst von Hwichan Choo und 3 weiteren Autor:innen.
Abstract (Originalsprache)
Jailbreaking is an AI attack method that manipulates prompts to elicit responses that violate AI policies. This study focuses on examining the impact of prompt engineering and the “Do Anything Now” (DAN) attack on commercial large language models (LLMs). For this, we gathered and generated a variety of prompts based on their type, purpose, and structure, inputted them into several commercial models, and analyzed the results to identify which components had the most significant impact on the success of jailbreak attacks. The findings revealed significant differences in jailbreak success rates based on prompt type and model configuration, with prompt engineering proving especially influential in bypassing restrictions. Additionally, specifying response formats or presenting questions as hypothetical scenarios was found to be particularly effective. This research aims to contribute to the development of techniques to prevent criminal activities involving artificial intelligence.
Zitieren
APA
Choo, H., Lee, C., Jeong, J., & Kim, K. (2024). Analysis of Jailbreak Prompt Vulnerabilities in Commercial LLM. Journal of Data Forensics Research, 1(1), 89-107. https://doi.org/10.12972/jdfr.2024.1.1.6
Vancouver
Choo H, Lee C, Jeong J, Kim K. Analysis of Jailbreak Prompt Vulnerabilities in Commercial LLM. Journal of Data Forensics Research. 2024;1(1):89-107. doi:10.12972/jdfr.2024.1.1.6
BibTeX
@article{Choo2024,
author = {Choo, Hwichan and Lee, Chanhwi and Jeong, Jaehoon and Kim, Kyungjong},
title = {Analysis of Jailbreak Prompt Vulnerabilities in Commercial LLM},
journal = {Journal of Data Forensics Research},
year = {2024},
volume = {1},
number = {1},
pages = {89--107},
doi = {10.12972/jdfr.2024.1.1.6}
}
EUVIMED