Whole slide images (WSIs) are essential in pathology diagnosis but cannot be directly applied to conventional classification networks due to their high resolutions. Although WSIs can be segmented into patches for analysis by neural networks, acquiring patch-level labels for this approach presents significant challenges. Consequently, multi-instance learning (MIL) is widely used for WSI classification, treating WSIs as bags and patches as instances. Current MIL methods often suffer from reduced accuracy due to the influence of easily recognizable patches within the WSIs. There is a pressing need for effective hard instance mining methods, as existing approaches struggle to identify hard instances most beneficial for enhancing MIL performance. Addressing this issue, we propose an adversarial multi-instance learning framework (AIM-MIL), which uses a learnable instance masking module (LIMM) to autonomously mask certain instances and a MIL module to aggregate the remaining instances. The interaction between these two modules facilitates the precise localization of hard instances and enhances learning from them. Additionally, we introduce gaussian loss and KL loss to improve the accuracy and complexity of the generated masks. Extensive experiments on Camelyon16 and BRACS datasets demonstrate that AIM-MIL significantly enhances performance. Furthermore, the framework can be integrated into existing attention-based MIL models, showcasing good adaptability and scalability.