The detection of IoT and internet traffic real-time zero-day multi-class attacks (novel or unseen attacks) are clearly explained in this research proposal. The proposed Stacking of four heterogeneous neural networks with a special type of Neural Network Autoencoder with ensemble Machine Learning Random Forest classifier delivers the best accuracy ratings and F1 Score of 0.999896 and 0.9998898 with the most minor loss function and the quickest execution times. Five heterogeneous neural networks with bagging machine learning make up this novel proposed stacking ensemble model. The highest scores were determined by comparing and evaluating the ensemble Random Forest Classifier with other ML classifiers, extreme Gradient Boosting, and Support Vector Machine are included. (XGB Boost), and the Naive Bayes. The applications of the main AI-deep learning models, ML classifiers, stacked deep learning models, and Stacked Ensemble Neural Network models with ML are then shown in this research analysis experimented over more than 20 lakhs of dataset instances on the realistic cyber and IoT datasets, which helps illuminate how different AI models are implemented for detecting zero-day attacks in network intrusion detection systems. Utilizing cutting-edge AI by implementing the proposed Stacking, Ensemble Stacking of DL, and ML Neural Networks with Feature Extractor significantly improves anomaly detection in identifying zero-day attacks. Therefore, it would effectively lessen attacks on IoT and cyber-security firms.