Hyperspectral anomaly detection, identifying targets that significantly differ from their surrounding background in an unsupervised manner, has recently favored autoencoder-based detectors for their satisfactory performance. However, these autoencoder-based approaches are predicated on the assumption that anomalies are harder to reconstruct than normal samples, an assumption that may not always true in practice. Additionally, processing anomalies and background together during reconstruction process limits the detector’s discriminative capability. To address these issues, this paper proposes a novel hyperspectral anomaly detection method based on unsupervised dynamic negative sampling contrastive learning. Firstly, the method incorporates overlapping spectral embeddings and multiscale interaction modules within a Transformer network to capture spectral features at varying granularities. Subsequently, a dynamic negative sampling strategy is employed to update the negative sample feature queue, reducing the module's sensitivity to data imbalance. Finally, the designed contrastive learning framework directly compares the similarity between pairs of spectral samples, enhancing the module's spectral discriminative capability and thus enabling anomaly detection in the feature space. Experimental results demonstrate the superiority of the proposed method in terms of comprehensive detection performance and background suppression.