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

This paper describes an architectural backdoor for vision-language model supply chains. A malicious provider embeds dormant, trigger-gated steering logic into model architecture or related executable artifacts. The modification is zero when the trigger is absent, preserving clean behavior, but shifts an intermediate representation toward an attacker-defined objective when activated. The authors report evaluations across multiple VLM families and tasks including visual question answering, text-to-image generation, retrieval, semantic response biasing, safety enforcement, and ranking fairness. The paper argues that auditing must inspect executable model logic, not only learned weights.