Abstract: In response to the challenges in the field of cybersecurity risk detection, such as the difficulty in pinpointing genuine attacks, low efficiency in risk assessment, judgment and disposal, and the high technical requirements for security personnel, a deep threat detection framework based on a SecurityGPT is proposed. This paper constructed a high-performing generative artificial intelligence large model tailored for the vertical domain of cybersecurity through corpus construction, model pre-training, instruction fine-tuning, and model inference acceleration. On this foundation, to further enhance the accuracy and detection efficiency of the model, a multi-dimensional collaborative research was conducted focusing on the integration of the security large model with traditional rule-based models and small-scale machine learning models. This initiative aimed to establish a tripartite deep threat detection architecture and tested in actual business environments. Experimental results show that this framework can ensure an average network risk detection rate of over 95% with a false positive rate below 5%, while significantly improving detection efficiency and reducing labor costs, demonstrating excellent application value.