Research map / Video / Video-language understanding
Breaking Multimodal LLM Safety via Video-Driven Prompting
Dong Wang, Xiangyu He, Xinqi Lyu, Bin Xiao
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 |
|---|---|---|---|
| repeat the same harmful image across frames to form a video (image-stack) instead of a single image (ablation) | FigStep · JailBreakV-28K, average over 5 models | ASR (%): 25.9 -> 28.6 | Improved |
| use diverse safety-proximal frames via bipartite matching instead of identical frames | SPTV vs FigStep (S) · JailBreakV-28K, average over 5 models | ASR (%): 28.6 -> 36.4 | Improved |
| stack identical frames of typographic image into video (ablation) | Typo / SD+Typo attacks · JailBreakV-28K, average over 5 models | ASR (%): SD+Typo 24.0 -> 22.7; Typo 12.5 -> 13.3 | Mixed |
| add perturbation-based VisualADV attack to video modality | VisualADV · JailBreakV-28K, average over 5 models | ASR (%): 6.7 -> 5.6 | No clear effect |
| image-stack unsafe video representation vs unsafe image (feature similarity to safe data) (ablation) | Qwen2-VL / Qwen2.5-VL · JailBreakV-28K | refusal probability | Improved |
| higher feature similarity to safe data (ablation) | Qwen2.5-VL image-stack · JailBreakV-28K 160 samples | correlation with log refusal probability: r = 0.83 | Improved |
| use video-aware system prompt (VSP) as defense | SPTV attack · Defense study | ASR (%): 38.8 -> 26.3 | Improved |
| use FigStep system prompt (FSD) as defense | SPTV attack · Defense study | ASR (%): 38.8 -> 35.6 | No clear effect |
| use video-aware system prompt (VSP) as defense | FigStep (S) attack · Defense study | ASR (%): 25.0 -> 0.0 | Improved |
| evaluate with GPT-4o-mini judge instead of LLaMA-Guard (ablation) | SPTV · Qwen2-VL-7B | ASR (%): 44.1 -> 49.4 | No clear effect |
| reduce text opacity of SPTV subtitle overlay on natural videos from 1.0 to 0.4 (ablation) | SPTV on natural videos · MSVD backgrounds, Qwen2.5-VL | ASR (%): 38.8 -> 33.1 | No clear effect |
Closest papers by meaning
- SafePTR: Token-Level Jailbreak Defense in Multimodal LLMs via Prune-then-Restore Mechanism (NeurIPS 2025 · similarity 0.92)
- Safe + Safe = Unsafe? Exploring How Safe Images Can Be Exploited to Jailbreak Large Vision-Language Models (NeurIPS 2025 · similarity 0.91)
- Playing the Fool: Jailbreaking LLMs and Multimodal LLMs with Out-of-Distribution Strategy (CVPR 2025 · similarity 0.90)
- Video-SafetyBench: A Benchmark for Safety Evaluation of Video LVLMs (NeurIPS 2025 · similarity 0.90)
- FigStep: Jailbreaking Large Vision-Language Models via Typographic Visual Prompts (AAAI 2025 · similarity 0.90)
- When Think-with-Image Meets Safety: What Determines Multimodal Jailbreak Robustness? (NeurIPS 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.
