arrow
Return

AWMT: Automatic Jailbreaking Attack Framework Utilizing Working-Memory Trees

delete2025-12-04
delete0
PRE
AI
张志强 (Zhiqiang Zhang)
J
Junjie Xu
李冰 (Bing Li)
Y
Yuankang Sun
H
Haimiao Mo
Y
Y.W Chen
DOI:10.1016/j.eswa.2025.130643delete
deleteOriginal
deleteOriginal request for help
deleteShare
deleteSave
Abstract

Abstract

En 中文
• AWMT framework uses tree structure and memory to boost attack efficiency. • Multi-prompt strategy improves jailbreak success rate and diversity. • Achieves 86% attack success on GPT-3.5-turbo, surpassing all baseline. • Jailbreak prompts are interpretable and transferable across various LLMs.

Journal

Expert Systems with Applications cover
Expert Systems with Applications
IF:
7.5
Papers:
2.9W
Citations:
10.2W

Organization

Z
Zhejiang University of Finance and Economics
Scholars:
363
Papers: 260
Citations: 25
S
Southeast University
Scholars:
2.0W
Papers: 8.3K
Citations: 480
S
Southern University of Science and Technology
Scholars:
5.2K
Papers: 2.1K
Citations: 34
researcher View more organizations