The integration of IoT and EdgeAI into cyber-physical systems (CPS) is transforming critical infrastructures, including industrial control systems, smart transportation, and smart grids by enabling scalable, real-time, and low-latency operations with enhanced connectivity, but also introduces new security challenges. MQTT, a lightweight messaging protocol widely used in IoT-enabled environments, enables efficient communication but lacks robust security features. Consequently, such critical infrastructures are vulnerable to various attacks such as resource exhaustion and communication disruptions, which can lead to system instability or failures. In this paper, we present a lightweight anomaly detection model integrated into resource-constrained MQTT brokers operating at the edge. These EdgeAI-based brokers function as standard MQTT brokers while simultaneously performing real-time traffic analysis to detect abnormal resource consumption. For training and evaluation, we construct synthetic datasets that emulate both benign and malicious MQTT traffic, with data categorized into three risk levels: normal, borderline, and anomalous. Our model achieves high detection accuracy and precision (up to 99%) while maintaining a compact memory footprint of approximately 20 KB, underscoring its suitability for deployment on the edge.