The proliferation of malicious users on social platforms poses significant financial and psychological threats, with activities ranging from scams to the dissemination of illicit content. Existing malicious user prediction comprises supervised and self-supervised learning methods. However, the former relies on extensive labeled malicious users for training, while the latter typically focuses on one form of malicious activity and depends heavily on manually crafted rules and features during pre-training. Moreover, existing pre-training methods fail to effectively capture the crucial repetitive and sporadic behavior patterns of malicious users. To address these limitations, we propose a \textbf{Ma}licious User Behavior \textbf{P}re-training framework (MaP) to build pre-trained behavior models. MaP integrates malicious pattern recognition with behavior consistency augmentation and local disruption augmentation strategies for contrastive learning to capture repetitive and sporadic malicious patterns, respectively. We instantiate MaP on a billion-level behavior pre-training scenario within an industry context. Both online and offline evaluations validate the superior performance of MaP in malicious user detection and classification.