🏆 Tong's paper on Pragamatic Jailbreak got Best Paper Award at Responsible Generative AI @ CVPR, Congrats!
June 13 🇺🇸 Nashville, TN — We are delighted to announce that PhD candidate Tong Liu and collaborators received the Best Paper Award at the 2nd Responsible Generative AI (ReGenAI) workshop, co‑located with CVPR 2025, for their work “Multimodal Pragmatic Jailbreak on Text‑to‑Image Models.”
The paper is the first to uncover pragmatic jailbreaks in diffusion‑based text‑to‑image (T2I) systems—cases where individually safe image and text prompts combine to bypass safety filters and trigger disallowed content. The team introduces a public benchmark spanning nine leading T2I models and shows violation rates up to 70 %, highlighting urgent gaps in current safety pipelines.
Authors: Tong Liu, Zhixin Lai, Jiawen Wang, Gengyuan Zhang, Shuo Chen, Philip Torr, Vera Demberg, Volker Tresp, Jindong Gu
👉 Project page, PDF, code & dataset
Congratulations to Tong and the entire team on this outstanding achievement!