Multivariate time series anomaly detection has great potential in various fields such as industrial systems, healthcare, and finance. The complex spatiotemporal dependencies in modern multivariate time series data make anomaly detection a challenging task. Existing deep learning approaches lack effective mechanisms to simultaneously capture temporal patterns and variable interactions, especially for modeling nonlinear dynamics in real-world data. On the other hand, these methods also face the risk of potential under-extraction of features. To address these challenges, we propose a GAN-based anomaly detection approach (MATS-GAN) with multiple attention mechanisms. MATS-GAN integrates complementary attention mechanisms within a bidirectional LSTM architecture to accurately capture complex spatiotemporal relationships. To overcome the feature extraction limitations, MATS-GAN introduce a label-guided mechanism that dynamically refines anomaly representations using confidence-based soft and hard labels. Experiments on three publicly available datasets demonstrate that MATS-GAN outperforms existing approaches, improving F1(K = 100) scores by up to 27.25 points and AUC values by over 46.18 points compared to alternative methods. These results confirm our method’s effectiveness in modeling the intricate dependencies in multivariate time series, thus enhancing anomaly detection performance.