paperJuly 2025Unreviewed

CAVGAN: Unifying Jailbreak and Defense of LLMs via Generative Adversarial Attacks on their Internal Representations

Xiaohu Li, Yunfeng Ning, Zepeng Bao, Mayi Xu, Jianhao Chen, Tieyun Qian

Annual Meeting of the Association for Computational Linguistics

Abstract

Security alignment enables the Large Language Model (LLM) to gain the protection against malicious queries, but various jailbreak attack methods reveal the vulnerability of this security mechanism. Previous studies have isolated LLM jailbreak attacks and defenses. We analyze the security protection mechanism of the LLM, and propose a framework that combines attack and defense. Our method is based on the linearly separable property of LLM intermediate layer embedding, as well as the essence of ja

Framework mappings

OWASP Top 10 for LLM Applications

LLM01Prompt Injection

MITRE ATLAS

AML.T0043Craft Adversarial Data
AML.T0054LLM Jailbreak

Suggested from the entry's categories.

Cite

@article{li2025cavgan,
  title = {{CAVGAN: Unifying Jailbreak and Defense of LLMs via Generative Adversarial Attacks on their Internal Representations}},
  author = {Xiaohu Li and Yunfeng Ning and Zepeng Bao and Mayi Xu and Jianhao Chen and Tieyun Qian},
  year = {2025},
  month = jul,
  journal = {Annual Meeting of the Association for Computational Linguistics},
  eprint = {2507.06043},
  archivePrefix = {arXiv},
  doi = {10.18653/v1/2025.findings-acl.346},
  url = {https://www.semanticscholar.org/paper/ff5db18341d678f2a9528dcd07ab5b63c031331c}
}