<p>The proliferation of hate speech on digital platforms has evolved from primarily text-based expressions to more sophisticated multimodal forms, combining offensive language with harmful images. With the advent of AI-generated content, this phenomenon has become more complex, making it increasingly difficult for conventional hate speech detection models to address it adequately. Adding to this challenge is the emergence of adversarial attacks, which deliberately manipulate inputs to mislead detection models, further exposing their vulnerabilities and reducing their effectiveness. This paper investigates the vulnerabilities of multimodal classifiers in detecting AI-generated multimodal hate speech when exposed to adversarial attacks and focuses on crafting adversarial examples that simultaneously exploit textual and visual components. Applying perturbations in the direction of the model’s loss function gradient demonstrates how minimal, imperceptible changes to both text and image inputs can deceive state-of-the-art classifiers. A novel technique, ADVattack, was introduced for generating adversarial examples using cross-model feature manipulation in this multimodal context and analyzing current models’ susceptibility to such attacks. The results indicate a significant gap in the robustness of existing multimodal hate speech detectors, highlighting the need for improved defense mechanisms against these sophisticated adversarial threats.</p>