paperJuly 2026Unreviewed

Architectural Backdoors in Vision-Language Model Supply Chains via Representation Steering

Maria Rosaria Briglia, Igor Maljkovic, Antonio Emanuele Cinà, Luca Oneto, Iacopo Masi, Fabio Roli

Abstract

Vision--Language Models (VLMs) are increasingly deployed through a model supply chain in which pretrained checkpoints, architecture definitions, text encoders, and exported computation graphs are distributed by third parties and reused across downstream services. This reuse model creates a security-critical trust boundary: VLM deployments inherit not only learned parameters but also executable behavior encoded in shared model artifacts. In this paper, we show that a malicious provider can exploi

Framework mappings

OWASP Top 10 for LLM Applications

LLM03Supply Chain
LLM04Data and Model Poisoning

MITRE ATLAS

AML.T0010AI Supply Chain Compromise
AML.T0020Poison Training Data

Suggested from the entry's categories.

Cite

@misc{briglia2026architectural,
  title = {{Architectural Backdoors in Vision-Language Model Supply Chains via Representation Steering}},
  author = {Maria Rosaria Briglia and Igor Maljkovic and Antonio Emanuele Cinà and Luca Oneto and Iacopo Masi and Fabio Roli},
  year = {2026},
  month = jul,
  eprint = {2607.25479},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2607.25479}
}