With the popularization of Internet applications, new APT attacks and encrypted traffic threats have grown exponentially, and the volume and complexity of network threats have risen. In addition, hackers have also been using ChatGPT, DeepSeek and other AI large model technology to rapidly and low-threshold generate high-adversarial attack threats, with strong bypass capabilities for traditional security equipment. Traditional regular rules, grammar and semantic, threat intelligence and other security solutions and technical measures are difficult to satisfy the requirements of real-time adversarial and accurate protection, thus artificial intelligence is widely used in the field of network security. This paper proposes a layered architecture based on the mixture of experts (MoE) to build a cyber security AI model framework covering threat detection, intelligent research and judgment, and attack traceability through studying and applying key technologies such as the principle of vertical field large model technology, threat detection, and intelligent operation. This design realizes model lightweight and scenario-based through reinforcement learning and knowledge distillation technology, achieves high accuracy in Web threat detection, intelligent threat research and judgment, 0 day vulnerability identification and other scenarios, and combines with security agent synergy mechanism to promote the cyber security defense to leap from reactive response to proactive prediction, provides a technical paradigm for the new generation of intelligent security system, improves the security capability of the organization.