Research map / Video / Video generation
RunawayEvil: Jailbreaking the Image-to-Video Generative Models
Songping Wang, Rufan Qian, Yueming Lyu, Qinglong Liu, Linzhuang Zou, Jie Qin, Songhua Liu, Caifeng Shan
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 |
|---|---|---|---|
| Use RunawayEvil instead of T2I jailbreak baselines (Sneaky, PGJ, DACA) | I2V jailbreak · COCO2017, 4 I2V models x 3 judges | ASR (%): average 87.6%, surpassing others by 58.5%-79% | Improved |
| Add Strategy Memory Bank (ablation) | RunawayEvil · COCO2017, Wan, Qwen-VL judge | ASR (%): 66.0% -> 71.0% | Improved |
| Add Strategy Customization Agent (RL) instead of random selection (ablation) | RunawayEvil · COCO2017, Wan, Qwen-VL judge | ASR (%): 75.0% -> 75.0% | No clear effect |
| Add Strategy Exploration Agent (ablation) | RunawayEvil · COCO2017, average over models and judges | ASR (%): 74.2% -> 87.2% | Improved |
| Attack both modalities jointly instead of text only (ablation) | RunawayEvil · COCO2017, average | ASR (%): 51.2% -> 87.2% | Improved |
| Attack both modalities jointly instead of image only (ablation) | RunawayEvil · COCO2017, average | ASR (%): 47.8% -> 87.2% | Improved |
| Coordinate text and image attacks instead of attacking each in isolation (ablation) | RunawayEvil · COCO2017, average | ASR (%): 79.7% -> 87.2% | Improved |
| Use adaptive steps with feedback instead of fixed 10 steps without feedback (ablation) | RunawayEvil · COCO2017, Wan, Qwen-VL judge | ASR (%): 68.0% -> 86.0% | Improved |
| Use 10 steps instead of 1 step without feedback (ablation) | RunawayEvil · COCO2017, Wan, Qwen-VL judge | ASR (%): 48.0% -> 68.0% | Improved |
| Add iteration penalty reward to success reward (ablation) | SCA reward · COCO2017, Wan, Qwen-VL judge | ASR (%): 80.0% -> 83.0% | Improved |
| Add text stealth reward (ablation) | SCA reward · COCO2017, Wan, Qwen-VL judge | ASR (%): 83.0% -> 84.0% | No clear effect |
| Add image stealth (LPIPS) reward (ablation) | SCA reward · COCO2017, Wan, Qwen-VL judge | ASR (%): 84.0% -> 86.0% | No clear effect |
Closest papers by meaning
- T2V-OptJail: Discrete Prompt Optimization for Text-to-Video Jailbreak Attacks (NeurIPS 2025 · similarity 0.90)
- MIDAS: Multi-Image Dispersion and Semantic Reconstruction for Jailbreaking MLLMs (ICLR 2026 · similarity 0.87)
- T2VSafetyBench: Evaluating the Safety of Text-to-Video Generative Models (NeurIPS 2024 · similarity 0.87)
- When the Prompt Becomes Visual: Vision-Centric Jailbreak Attacks for Large Image Editing Models (ICML 2026 · similarity 0.87)
- I2VGuard: Safeguarding Images against Misuse in Diffusion-based Image-to-Video Models (CVPR 2025 · similarity 0.87)
- ColJailBreak: Collaborative Generation and Editing for Jailbreaking Text-to-Image Deep Generation (NeurIPS 2024 · similarity 0.86)
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.
