How to fool AI that sees in infrared and visible light

Vision-language models work well on multimodal tasks but haven't been tested against adversarial attacks in visible-infrared scenarios—a gap that matters because these sensors are used in real-world systems. Researchers built CFGPatch, an adversarial patch that uses curved fractal geometry and spiral texture patterns to fool both visible and infrared inputs simultaneously. The patches transferred effectively across different downstream tasks like image captioning and visual QA, suggesting these models share exploitable vulnerabilities regardless of their final application.