Recommender systems play a pivotal role in tackling information overload but are often faced with problems of data sparsity and cold start. Multi-behavior Sequential Recommendation (MBSR) tackles these challenges by taking diverse user-item interaction behaviors into account. However, existing MBSR models are often challenged by the imbalance issue in multi-behavior data. Traditional data augmentation techniques for MBSR are insufficient as they concentrate on single-behavior data and fail to account for the interplay between behaviors and the dynamics of sequences. To address these shortcomings, we propose a diffusion based data augmentation framework for multi-behavior sequential recommendation (DiffAMB). We utilize diffusion models and contrastive learning to generate high quality data for MBSR models. This framework enables the augmentation of multiple-behavior datasets while preserving the logical consistency of user sequences. Our experiments on real-world datasets confirm the effectiveness of our approach.