The proliferation of resource-constrained Internet-of-Things (IoT) and Industrial IoT (IIoT) devices has intensified the need for local, intelligent cybersecurity solutions that do not rely on cloud connectivity. This paper presents a TinyML-based Intrusion Detection System (IDS) framework designed for embedded edge controllers. Two machine-learning IDS models - a Convolutional Neural Network (CNN) and a Random Forest (RF) - were developed and evaluated on the TON-IoT dataset to identify malicious network traffic. Host-side evaluation established baseline performance, while the float-precision RF model was implemented as firmware and validated using hardware-in-the-loop emulation in the Renode RISC-V environment. The resulting binary occupies only 1.6 MB of Flash memory, achieving above 99% detection accuracy with 448 μs average inference latency. These results demonstrate that fully embedded, real-time intrusion detection is feasible within TinyML memory budgets, supporting IEC-62443 compliant anomaly monitoring for industrial devices.