Multimodal Pragmatic Jailbreak on Text-to-image Models

Tong Liu¹² Zhixin Lai⁵ Jiawen Wang¹ Gengyuan Zhang¹² Shuo Chen¹²
Philip Torr⁶ Vera Demberg³⁴ Volker Tresp¹² Jindong Gu⁶

¹LMU Munich, Germany ²Munich Center for Machine Learning, Germany
³ Saarland University, Germany ⁴Max Planck Institute for Informatics, Germany
⁵Cornell University, USA ⁶University of Oxford, UK

ACL 2025; Best Paper Award 🏆 at ReGenAI@CVPR 2025

[Paper] [Dataset] [Code]

Summary: we first propose that AI safety can be 𝐦𝐮𝐥𝐭𝐢𝐦𝐨𝐝𝐚𝐥!

BibTeX

 @inproceedings{liu-etal-2025-multimodal-pragmatic,

    title = "Multimodal Pragmatic Jailbreak on Text-to-image Models",

    author = "Liu, Tong and Lai, Zhixin  and Wang, Jiawen and Zhang, Gengyuan and Chen, Shuo and Torr, Philip and Demberg, Vera and Tresp, Volker and Gu, Jindong",

    editor = "Che, Wanxiang and Nabende, Joyce and Shutova, Ekaterina and Pilehvar, Mohammad Taher",

    booktitle = "Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",

    month = jul,

    year = "2025",

    address = "Vienna, Austria",

    publisher = "Association for Computational Linguistics",

    url = "https://aclanthology.org/2025.acl-long.234/",

    doi = "10.18653/v1/2025.acl-long.234",

    pages = "4681--4720",

    ISBN = "979-8-89176-251-0"

    }

License

The MPUP dataset is released under the CC BY-4.0 License.