Understanding and Defending VLM Jailbreaks via Jailbreak-Related Representation Shift
Published in NeurIPS, 2026
Recommended citation: Wei, Z., Li, Q., Ruan, J., Qin, Z., Wen, L., Qin, R., Wang, Q., Liu, D., & Shen, W. Understanding and Defending VLM Jailbreaks via Jailbreak-Related Representation Shift. In NeurIPS 2026. https://arxiv.org/pdf/2603.17372
Abstract. This paper studies how visual modality can induce jailbreak-related representation shifts in vision-language models and proposes a defense that removes the jailbreak-related shift at inference time.
Authors: Zhihua Wei*, Qiang Li*, Jian Ruan, Zhenxin Qin, Leilei Wen, Ruiyang Qin, Qingzhuo Wang, Dongrui Liu, Wen Shen†.
* Equal contribution; † Corresponding author.