Many of these critical infrastructures have incorporated AI-based systems, for instance, in precision irrigation; therefore, securing communication networks becomes ever more critical. Traditional anomaly detection approaches in cyber-physical systems are ineffective in dealing with the dynamic nature of IoT networks to avoid vulnerabilities such as data spoofing and denial-of-service (DoS) attacks. This research uses an artificial intelligence-based methodology to ensure the security of the communication that is to be established within a cyber-physical system (CPS) by adopting recurrent neural networks (RNNs), random forests (RFs), autoencoders, and graph neural networks (GNNs). The training dataset was comprised of IoT sensor data, communication logs, and simulated cyberattacks, both normal as well as anomalous communication patterns. The respective models were checked against real-time anomaly detection and identification of affected devices. By analysing the results, the accuracy level for RNN came out to be at the top, with a value of 98.89% with minimal latency and high throughput, which makes this model highly suitable for real-time applications. RF, Autoencoders, and GNN followed this at accuracies of 94.5%, 89.90%, and 87.6%, respectively. The proposed models were successful in flagging anomalies and consequently initiating the isolation of affected devices without the need for human intervention, thereby ensuring the integrity of the system. This approach dramatically applies to securing IoT-enabled systems in sectors like intelligent agriculture, wherein real-time anomaly detection and rapid response are necessary to prevent system disruptions. The research shows the probable capabilities of machine learning (ML) in improving the security of critical infrastructures through timely automatic responses to emerging threats.