A recent study proposes a novel attack method, named DURA, targeting Vision-Language-Action (VLA) models. This method utilizes a diffusion model to generate natural adversarial patches, allowing the manipulation of robots to perform specific actions. DURA supports both white-box and black-box attack settings, requiring only the predicted actions of the victim model. Experimental results demonstrate that DURA outperforms existing methods in both simulation and the physical world. This attack method exposes a safety risk for VLA models in the physical world, calling for stronger defenses. Blogger's Review: This article presents an innovative attack method, highlighting the security risks of VLA models and emphasizing the need for secure deployment of AI models in the physical world. It urges researchers to strengthen defense measures, ensuring the safety and reliability of AI systems.