This paper introduces a hybrid cybersecurity framework integrating artificial intelligence and mathematical modeling for real-time intrusion detection. Deep learning techniques (CNN-LSTM) are coupled with a stochastic threat-behavior model based on Markov decision processes (MDPs) to predict and classify cyberattacks. The mathematical model quantifies the likelihood of malicious transitions in network states, while the AI component enhances pattern recognition and adaptive learning. Experimental validation using CICIDS2017 dataset shows a detection accuracy of 98.7% and reduced false positives. The framework’s response mechanism is optimized using dynamic threshold adjustment derived from probabilistic inference. This dual-model approach ensures both analytical transparency and autonomous defense capabilities, making it suitable for complex, evolving cyber environments.