Visual Adversarial Examples Jailbreak Aligned Large Language Models

Abstract

Shows adversarial images can jailbreak multimodal LLMs that are robust to text-only attacks, bypassing alignment through the visual channel.

Categories

Framework mappings

@inproceedings{qi2024visual,
  title = {{Visual Adversarial Examples Jailbreak Aligned Large Language Models}},
  author = {Xiangyu Qi and Kaixuan Huang and Ashwinee Panda and Peter Henderson and Mengdi Wang and Prateek Mittal},
  year = {2024},
  booktitle = {AAAI 2024},
  eprint = {2306.13213},
  archivePrefix = {arXiv},
  doi = {10.1609/aaai.v38i19.30150},
  url = {https://arxiv.org/abs/2306.13213}
}