Rapid landslide risk zoning is a critical task for disaster prevention and emergency management, particularly in mountainous regions where slope stability is strongly influenced by spatially heterogeneous geotechnical parameters and complex environmental loading. Although high-fidelity numerical models, such as the smoothed particle hydrodynamics (SPH) [1] method for simulating large-deformation landslide processes, are capable of capturing these nonlinear behaviors, their computational cost often renders them impractical for large-scale or near–real-time risk assessment. To address this challenge, this study proposes a rapid landslide risk zoning approach using a gradient-enhanced active learning Kriging framework. The proposed method incorporates gradient information available from the surrogate model into an adaptive Kriging framework to improve prediction accuracy in the vicinity of the failure boundary. An active learning strategy is employed to iteratively enrich the training set by selectively adding samples in regions with high epistemic uncertainty. Compared with conventional Kriging-based reliability or risk assessment approaches, the proposed framework significantly reduces the number of expensive model evaluations required to achieve a prescribed accuracy level. Based on the constructed surrogate, the probabilistic landslide risk of an individual slope is efficiently identified and quantified, and the associated influence area is further zoned according to the spatial distribution of failure probability under uncertain geotechnical conditions. The performance of the proposed method is demonstrated through representative slope stability examples involving spatially variable soil properties, where large-deformation landslide responses are simulated using SPH method. The results indicate that accurate estimates of failure probability and risk zoning boundaries can be obtained with substantially lower computational cost than traditional Monte Carlo simulation and standard adaptive Kriging methods. Owing to its efficiency and robustness, the proposed approach is well suited for applications requiring repeated analyses, such as rainfall-induced landslide risk assessment and rapid hazard mapping.