Research map / Adversarial robustness and security / Attacks on text and graphs
Toward Universal and Transferable Jailbreak Attacks on Vision-Language Models
Kaiyuan Cui, Yige Li, Yutao Wu, Xingjun Ma, Sarah Monazam Erfani, Christopher Leckie, Hanxun Huang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the optimised jailbreak image (vs text-only input) (ablation) | UltraBreak · SafeBench black-box average | ASR (%): 40.79 -> 71.83 | Improved |
| add the optimised jailbreak image (vs text-only input) (ablation) | UltraBreak · AdvBench black-box average | ASR (%): 25.48 -> 57.64 | Improved |
| add random transformations and TV constraints on the image (ablation) | UltraBreak · SafeBench black-box average | ASR (%): 51.99 -> 71.83 | Improved |
| add random transformations and TV constraints on the image (ablation) | UltraBreak · AdvBench black-box average | ASR (%): 29.86 -> 57.64 | Improved |
| add random transformations and TV constraints on the image (ablation) | UltraBreak · SafeBench, white-box Qwen2-VL surrogate | ASR (%): 89.84 -> 81.59 | Hurt |
| replace cross-entropy with the attention semantic loss (ablation) | UltraBreak · SafeBench black-box average | ASR (%): 55.80 -> 71.83 | Improved |
| replace cross-entropy with the attention semantic loss (ablation) | UltraBreak · AdvBench black-box average | ASR (%): 40.15 -> 57.64 | Improved |
| add attention weighting to the semantic loss (ablation) | UltraBreak · SafeBench black-box average | ASR (%): 57.54 -> 71.83 | Improved |
| add attention weighting to the semantic loss (ablation) | UltraBreak · AdvBench black-box average | ASR (%): 41.83 -> 57.64 | Improved |
| increase TV loss weight from 0 to 0.2 (ablation) | UltraBreak · SafeBench, Qwen-VL and GLM-4.1V average | ASR (%): 65.88 -> 72.38 | Improved |
| increase TV loss weight from 0.5 to 1 (ablation) | UltraBreak · SafeBench, Qwen-VL and GLM-4.1V average | ASR (%): 69.37 -> 63.18 | Hurt |
| change the affirming phrase in TPG (ablation) | UltraBreak · SafeBench, GLM-4.1V-9B-Thinking | ASR (%): 66.03 / 72.06 / 66.98 | No clear effect |
Closest papers by meaning
- Failures to Find Transferable Image Jailbreaks Between Vision-Language Models (ICLR 2025 · similarity 0.93)
- Jailbreak in pieces: Compositional Adversarial Attacks on Multi-Modal Language Models (ICLR 2024 · similarity 0.92)
- Steering Away from Harm: An Adaptive Approach to Defending Vision Language Model Against Jailbreaks (CVPR 2025 · similarity 0.91)
- JailBound: Jailbreaking Internal Safety Boundaries of Vision-Language Models (NeurIPS 2025 · similarity 0.91)
- VERA-V: Variational Inference Framework for Jailbreaking Vision-Language Models (ICML 2026 · similarity 0.91)
- Jailbreaking Vision-Language Models via Dissonance-Guided Suffix Optimization and Image-Phrase Injection (CVPR 2026 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
