Researchers have found that vision language models (VLMs) can be easily manipulated by adding small, imperceptible changes to visual inputs. This makes VLMs vulnerable to attacks that disrupt their ability to interpret images and generate text descriptions. The study proposes a new method for generating these adversarial examples, which is more efficient than previous approaches but still effective in altering the model's output. The findings have implications for the security of multimodal AI systems used in real-world applications. --- Why it matters: This matters because VLMs are widely used in safety-critical applications such as self-driving cars and medical imaging analysis. If these models can be easily manipulated, it could lead to serious consequences, including incorrect diagnoses or accidents on the road. Source: https://arxiv.org/abs/2608.18938