Architectural Backdoors Threaten Vision-Language Model Supply Chains

A recent arXiv preprint (2607.25479) indicates that harmful actors can insert architectural backdoors into Vision-Language Model (VLM) supply chains through a method known as representation steering. This technique incorporates hidden steering logic into the model's architecture using a trigger-gated additive alteration of an intermediate representation, all while avoiding the need to poison training data, manage downstream fine-tuning, or change prompts during deployment. In the absence of the trigger, the alteration effectively becomes null. This tactic takes advantage of the trust boundary present in VLM implementations that rely on executable behaviors from model artifacts shared by external sources.