ComciJailbreak introduces a comic-based jailbreak dataset to evaluate whether MLLMs uphold safety policy when harmful goals are embedded in visual narrative.
To create ComicJailbreak, run the following code:
pip install -r requirements.txt
python create_dataset.py --type articleWe have released the ComicJailbreak dataset, and will provide the inference and evaluation code when they are ready. Stay tuned...
If you find this work useful in your research, please cite the following paper:
@article{tan2026structured,
title={Structured Visual Narratives Undermine Safety Alignment in Multimodal Large Language Models},
author={Tan, Rui Yang and Hu, Yujia and Lee, Roy Ka-Wei},
journal={arXiv preprint arXiv:2603.21697},
year={2026}
}