Con Instruction: Universal Jailbreaking of Multimodal Large Language Models via Non-Textual Modalities

This work demonstrates how multimodal large language models can be universally jailbroken through carefully crafted non-textual inputs, highlighting critical security concerns in multimodal AI systems.

Recommended citation: J Geng, TT Tran, P Nakov, I Gurevych. (2025). "Con Instruction: Universal Jailbreaking of Multimodal Large Language Models via Non-Textual Modalities." ACL 2025 Main.
Download Paper