New research from FIU shows that some visual-language AI models have become particularly susceptible to image-based hacks.