Deep neural networks (DNNs) have achieved impressive results in image classification but remain highly vulnerable to adversarial attacks, threatening safety-critical applications such as autonomous driving. Recent detection approaches based on mutual information (MI) show promise in building robust latent spaces with strong generalization. However, their discrimination mechanisms primarily rely on image feature distributions and lack modeling and guidance of semantic information, resulting in insufficient discriminability of latent representations. To address this, this paper proposes an adversarial example detection method based on cross-modal semantic consistency. First, a dual-stream encoder is employed to encode both the original image and its noisy version, enhancing the robustness of latent representations to minor perturbations. Subsequently, a structured and discriminative latent space is constructed through multi-level mutual information maximization (MIMax) and prior distribution matching (PDM). Building on this, a pre-trained CLIP model is introduced, whose text label embeddings contain rich semantic information. These text embeddings serve as semantic priors to guide image representations to cluster towards the correct semantics, effectively improving the discriminative power of the latent representations. Finally, a lightweight fully connected network projects the latent representations into a semantic feature space consistent with the target model’s output space. An adaptive threshold strategy is designed based on similarity distributions and model outputs to identify adversarial examples. Experimental results on MNIST, CIFAR-10, and ImageNet datasets show that the proposed method outperforms multiple baselines, demonstrating superior detection performance.