Cloaked in authority: a practical black-box jailbreak of MLLMs exploiting visual trust bias
Dressed in Authority: Practical Black-Box Jailbreaking of MLLMs via Visual Trust Bias
Researchers propose a black-box jailbreak framework for multimodal large language models (MLLMs), Visual Trust Bias Attack (VTBA), which exploits the model's intrinsic trust bias toward authoritative visual features to bypass safety refusals, achieving consistently high attack success rates on multiple strong MLLMs. Unlike prior methods that attack by injecting fragile adversarial perturbations or synthesizing toxic visual cues, VTBA acts directly on naturally occurring harmful image-text pairs, and the attack is transferable and stealthy. Methodologically, VTBA builds a library of authoritative templates, matches the most relevant template to each input image-text pair, and embeds the harmful intent into the selected layout, using trustworthy visual cues to override refusals while maintaining naturalness. The authors also construct a dataset of 2,500 natural image-text pairs for evaluation, and state that the results reveal a severe and previously understudied weakness in current multimodal safety alignment.