Internet of Things (IoT) ecosystems has been associated with serious cybersecurity threats due to the heterogeneous nature, dynamism in connectivity, and poor computing abilities. The present paper presents an AI-based early warning system, which is strong in detecting and predicting cyber threats in an IoT environment, with a high accuracy and low latency. The framework is a combination of sequential learning algorithms, i.e., BiLSTM and GRU, to learn temporal anomalies of network traffic, and Graph Neural Networks (GNN) to learn complex interactions among devices. A hybrid ensemble of Autoencoder based anomaly detectors and XGBoost based classification can achieve an improved predictive performance and a reduced false positives. It has a distributed monitoring and edge intelligence layer to mitigate threats in real time by processing real time data. The experimental results indicate that the proposed system is more accurate in detecting and has a shorter response time and higher system robustness and hence can be applied to scalable and security-critical internet of things applications in other domains.