Saved in:
Bibliographic Details
Main Authors: Li, Chongxin, Wang, Hanzhang, Duan, Lian
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2603.14219
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866912967005569024
author Li, Chongxin
Wang, Hanzhang
Duan, Lian
author_facet Li, Chongxin
Wang, Hanzhang
Duan, Lian
contents Safety prompts constitute an interpretable layer of defense against jailbreak attacks in vision-language models (VLMs); however, their efficacy is constrained by the models' latent structural responsiveness. We observe that such prompts consistently engage a sparse set of parameters that remain largely quiescent during benign use. This finding motivates the Safety Subnetwork Hypothesis: VLMs embed structurally distinct pathways capable of enforcing safety, but these pathways remain dormant without explicit stimulation. To expose and amplify these pathways, we introduce Safety-Potential Pruning, a one-shot pruning framework that amplifies safety-relevant activations by removing weights that are less responsive to safety prompts without additional retraining. Across three representative VLM architectures and three jailbreak benchmarks, our method reduces attack success rates by up to 22% relative to prompting alone, all while maintaining strong benign performance. These findings frame pruning not only as a model compression technique, but as a structural intervention to emerge alignment-relevant subnets, offering a new path to robust jailbreak resistance.
format Preprint
id arxiv_https___arxiv_org_abs_2603_14219
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Safety-Potential Pruning for Enhancing Safety Prompts Against VLM Jailbreaking Without Retraining
Li, Chongxin
Wang, Hanzhang
Duan, Lian
Computer Vision and Pattern Recognition
Safety prompts constitute an interpretable layer of defense against jailbreak attacks in vision-language models (VLMs); however, their efficacy is constrained by the models' latent structural responsiveness. We observe that such prompts consistently engage a sparse set of parameters that remain largely quiescent during benign use. This finding motivates the Safety Subnetwork Hypothesis: VLMs embed structurally distinct pathways capable of enforcing safety, but these pathways remain dormant without explicit stimulation. To expose and amplify these pathways, we introduce Safety-Potential Pruning, a one-shot pruning framework that amplifies safety-relevant activations by removing weights that are less responsive to safety prompts without additional retraining. Across three representative VLM architectures and three jailbreak benchmarks, our method reduces attack success rates by up to 22% relative to prompting alone, all while maintaining strong benign performance. These findings frame pruning not only as a model compression technique, but as a structural intervention to emerge alignment-relevant subnets, offering a new path to robust jailbreak resistance.
title Safety-Potential Pruning for Enhancing Safety Prompts Against VLM Jailbreaking Without Retraining
topic Computer Vision and Pattern Recognition
url https://arxiv.org/abs/2603.14219