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.1

Adversarial vulnerabilities in vision-language models pose a critical challenge to the reliability of large language systems, where typographic manipulations and adversarial perturbations can effectively bypass language model defenses. We introduce Attack as Defense (AsD), the first approach to proactively defend at the cross-modality level, embedding protective perturbations in vision to disrupt attacks before they propagate to the language model. By leveraging the semantic alignment between vision and language, AsD enhances adversarial robustness through model perturbations and system-level prompting. Unlike prior work that focuses on text-stage defenses, our method integrates visual defenses to reinforce prompt-based protections, mitigating jailbreaking attacks across benchmarks. Experiments on the LLaVA-1.5 show that AsD reduces attack success rates from 56.7% to 12.6% for typographic attacks and from 89.0% to 47.5% for adversarial perturbations. Further analysis reveals that the key bottleneck in vision-language security lies not in isolated model vulnerabilities, but in cross-modal interactions, where adversarial cues in the vision model fail to consistently activate the defense mechanisms of the language model.