Paper

Understanding and Defending VLM Jailbreaks via Jailbreak-Related Representation Shift

Large vision-language models (VLMs) often exhibit weakened safety alignment with the integration of the visual modality. Even when text prompts contain explicit harmful intent, adding an image can substantially increase jailbreak success rates. In this paper, we observe that VLMs can clearly distinguish benign inputs from harmful ones in their representation space. Moreover, even among harmful inputs, jailbreak sampl · Authors: Zhihua Wei, Qiang Li, Jian Ruan, Zhenxin Qin, Leilei Wen, Dongrui Liu… · Published 2026-03-18 · arXiv:2603.17372

Published
2026-03-18
Authors
7
arXiv
2603.17372
Category
cs.CV

Abstract

arXiv:2603.17372

Zhihua Wei, Qiang Li, Jian Ruan, Zhenxin Qin, Leilei Wen, Dongrui Liu, Wen Shen

Large vision-language models (VLMs) often exhibit weakened safety alignment with the integration of the visual modality. Even when text prompts contain explicit harmful intent, adding an image can substantially increase jailbreak success rates. In this paper, we observe that VLMs can clearly distinguish benign inputs from harmful ones in their representation space. Moreover, even among harmful inputs, jailbreak samples form a distinct internal state that is separable from refusal samples. These observations suggest that jailbreaks do not arise from a failure to recognize harmful intent. Instead, the visual modality shifts representations toward a specific jailbreak state, thereby leading to a failure to trigger refusal. To quantify this transition, we identify a jailbreak direction and define the jailbreak-related shift as the component of the image-induced representation shift along this direction. Our analysis shows that the jailbreak-related shift reliably characterizes jailbreak behavior, providing a unified explanation for diverse jailbreak scenarios. Finally, we propose a defense method that enhances VLM safety by removing the jailbreak-related shift (JRS-Rem) at inference time. Experiments show that JRS-Rem provides strong defense across multiple scenarios while preserving performance on benign tasks.

Related papers

22

  • Tongji University
  • University of California at Merced
  • Westlake University
  • Zhejiang University
  • King’s College London
  • Tongji University
  • University of Luxembourg
  • Beijing Institute of Aerospace Control Devices
  • Manhattan University
  • Ryazan State Radio Engineering University
  • Tongji University
  • National University of Singapore
  • Shanghai Artificial Intelligence Laboratory
  • Shanghai Innovation Institute
  • Shanghai Jiao Tong University
  • University of Cambridge
  • McGill University
  • The Hong Kong Polytechnic University
  • Tongji University
  • Tongji University
  • Xiamen University