Changing Font Colors Can Hijack AI Reasoning

A recent study reveals that visual styling like font color, brightness, and contrast can influence vision language models (VLMs) to misinterpret text semantics, affecting tasks like sentiment analysis and question answering. This exposes a vulnerability in VLMs that could impact their robustness and safety.