Increased concerns about hardware security have been raised by the high rate of agentic AI infrastructure development and the growing complexity of integrated circuits (ICs). The most susceptible to hardware Trojans are Network-on-Chip (NoC) architectures, which are essential to contemporary system-on-chip design. Such architectures should be made trustworthy by having powerful and intelligent detection systems. This paper solves this problem by implementing machine learning-based anomaly detector methods on an 8x8 mesh-based NoC. The Garnet 2.0 cycle-accurate simulator is used to generate communication traffic, which produces detailed metrics (packet latency, hop count, and buffer utilization) about the traffic. Advanced machine learning classifiers that are able to differentiate between normal behavior and anomalies caused by Trojans are trained using these features. The models have been optimized on high-dimensional traffic data, and focus on accuracy, recall, and the overall classification accuracy and with a minimum of false positives and false negatives. The findings of the experiment show that learning-based models can identify the activity of a Trojan with a high degree of reliability, and the detection accuracy of the model is greatly improved over the baseline statistical models. The suggested model emphasizes the incorporation of AI-driven security into the agentic computing infrastructures and establishes the level of resilience to semiconductor supply chains, and guarantees the integrity of the future NoC-based ICs.