The spread of infectious diseases, rumors, and harmful speech in networks can result in substantial losses, underscoring the significance of studying how to suppress such hazardous events. However, previous studies often assume full knowledge of the network structure, which is often not the case in real-world scenarios. In this paper, we address the challenge of controlling the propagation of hazardous events by removing nodes when the network structure is unknown. To tackle this problem, we propose a hierarchical reinforcement learning method that drastically reduces the action space, making the problem feasible to solve. Simulation experiments demonstrate the superiority of our method over the baseline methods. Remarkably, even though the baseline methods possess extensive knowledge of the network structure, while our method has no prior information about it, our approach still achieves better results.