As network architectures become increasingly dynamic and complex, traditional static defense mechanisms—relying on predefined rules—are proving inadequate in coping with evolving and unknown security threats. Inspired by the regulatory logic of the human nervous system, particularly its feed-forward and feedback mechanisms, conditioned reflexes, and self-adaptive learning, this paper proposes a neural-inspired security control model that integrates predictive regulation with adaptive optimization. The model employs a multi-hypersphere modeling approach to characterize multiple normal operating states of the system, simulating the nervous system's ability to flexibly reset regulatory baselines under different conditions. This enables accurate anomaly detection and feed-forward signal generation. In the feedback loop, a reward mechanism based on potential functions guides the system in progressively refining its countermeasure strategies, enhancing adaptability to unknown attacks. Experiments were conducted on a simulated Spring Boot based Web service platform, covering two normal states and seven categories of abnormal states. All experimental data were collected from the actual system runtime. Results show that the model achieves 98.00% accuracy in anomaly detection and 98.29% correctness in feed-forward signal generation, with all abnormal scenarios successfully regulated within a limited number of steps. Furthermore, comparative analysis demonstrates that the proposed model significantly outperforms standard anomaly detection baselines, such as OC-SVM and Isolation Forest, in both detection accuracy and regulatory capability. These findings validate the model's capabilities in self-learning, generalization, and practical application in intelligent security control.