ComicJailbreak

ComicJailbreak introduces a comic-based jailbreak dataset to evaluate whether MLLMs uphold safety policy when harmful goals are embedded in visual narrative.

Dataset

Dataset Creation

To create ComicJailbreak, run the following code (we are using uv for the environment setup):

uv sync

uv run python create_dataset.py --type article

• You can extend the dataset by creating a csv file that looks like this:

Article	Speech	Instruction	Message	Code
Your prompt	...	...	...	...

For paraphrasing, run the following code:

uv run python paraphrasing --goals <your_path.csv> --type all

Experiments

If you were to use the API inferences, please include .env file to store your API keys:

OPENROUTER_API=<YOUR_API_KEY>

We have included the prompts to perform the attack and defenses, packaged into shell scripts:

# For ComicJailbreak attack experiments
bash attack.sh              # Local inference
bash openrouter_attack.sh   # API inference (we are using OpenRouter as the provider)

# For defenses against ComicJailbreak
bash defense.sh             # Local inference
bash openrouter_defense.sh  # API inference (we are using OpenRouter as the provider)

# Script for evaluation
bash eval.sh                # Local inference

Citation

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}
}