Due to the increasing frequency and complexity of network attacks, traditional security measures are no longer able to cope with these refined threats. Therefore, in recent years, people have begun to advocate the use of the deep reinforcement learning (DRL) algorithm to enhance the responsiveness of network security. The article demonstrates the advantages of the DRL algorithm in different network threat scenarios through multiple experiments. The DRL algorithm achieved a 95% accuracy and an F1 score of 0.94 in benchmark attack detection and demonstrated good adaptability and resilience in zero-day attack and adversarial sample detection experiments. Finally, in the system performance stress test, as the network traffic increases, the DRL processing speed decreases, while the CPU and memory usage significantly increase, which also demonstrates the potential limitations of the DRL algorithm in practical applications.