The increasing integration of Internet of Things (IoT) devices in critical domains such as healthcare, smart homes, and industrial systems has introduced new cybersecurity challenges. This paper proposes a hybrid deep learning framework that combines a Long Short-Term Memory (LSTM)-based autoencoder with a Transformer layer for adaptive sequence modeling and anomaly detection in IoT networks. The model leverages LSTM encoders to capture local temporal dependencies and multi-head self-attention to model global context, enabling effective detection of anomalous behavior across time. Experimental evaluations on the CICIoMT2024 dataset demonstrate improved detection performance across various attack types, including DDoS, spoofing, and reconnaissance, achieving high accuracy and reduced false positive rates. The proposed architecture outperforms traditional LSTM and standalone Transformer baselines, making it suitable for real-time IoT threat detection applications.