Cross-Modal Backdoors in Multimodal Large Language Models

Abstract

Developers increasingly construct multimodal large language models (MLLMs) by assembling pretrained components,introducing supply-chain attack surfaces.Existing security research primarily focuses on poisoning backbones such as encoders or large language models (LLMs),while the security risks of lightweight connectors remain unexplored.In this work,we propose a novel cross-modal backdoor attack that exploits this overlooked vulnerability.By poisoning only the connector using a single seed sample

Categories

Framework mappings

Suggested from the entry's categories.

@misc{wang2026crossmodal,
  title = {{Cross-Modal Backdoors in Multimodal Large Language Models}},
  author = {Runhe Wang and Li Bai and Haibo Hu and Songze Li},
  year = {2026},
  month = may,
  eprint = {2605.07490},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2605.07490}
}