Imperceptible Visual Prompt Injection Threatens Commercial Multimodal LLMs

A recent study published on arXiv (ID: 2603.29418) presents a novel technique for covert visual prompt injection aimed at commercial multimodal large language models (MLLMs). Unlike traditional methods that depend on visible text or images, this strategy discreetly embeds harmful instructions within visual content. It employs a bounded text overlay for semantic direction while iteratively fine-tuning an undetectable visual alteration to synchronize the feature representation of the targeted image with both harmful visual and textual objectives at varying levels of detail. The visual target manifests as a text-rendered image, and the attack unfolds in stages. This research reveals a critical security flaw in MLLMs, which are increasingly utilized in practical scenarios. This paper is an updated version (replace-cross) of the initial submission, reflecting ongoing improvements and emphasizing the urgent need for strong defenses against such covert threats in AI systems.