The diverse and resource-constrained nature of Internet of Things (IoT) systems presents significant challenges for effective botnet detection. A novel hybrid deep learning approach is introduced by combining Gated Recurrent Units (GRUs) with Convolutional Neural Networks (CNNs). In this framework, CNNs are utilized to extract spatial information from network traffic, while GRUs capture temporal dependencies to enhance the detection of passive malicious activities. Using NF-UNSW-NB15-v2 and DNN-Edge IIoT datasets, performance is compared against an alternative GRU-LSTM hybrid model. With an accuracy of 100%, the CNN-GRU model attained precision scores between 0.99 and 1.00, recall of 1.00, and F1 scores also ranging from 0.99 to 1.00. Detection performance on the DNN-EdgeIIoT dataset was exact precision and recall value-based perfect. Whereas the GRU-LSTM model attained 99.10%, the alternative hybrid model recorded 99.35% accuracy. These results show the adaptability and efficiency of the CNN-GRU hybrid model in many IoT attack scenarios, so stressing the advantage of combining spatial and temporal learning mechanisms. Particularly in view of growing cybersecurity issues in IoT environments, the proposed architecture provides a strong basis for actual implementation of hybrid detection systems based on deep learning.