Insider threats are becoming one of the most serious challenges in cybersecurity, causing heavy financial and operational damage to organizations. Traditional rule-based detection methods often fail because they cannot recognize complex or new patterns of malicious behavior. To address this gap, we propose a hybrid insider threat detection model that combines Long Short-Term Memory (LSTM) autoencoders with both temporal activity logs and psychometric (behavioral) features, using the CERT CMU Insider Threat Dataset v6.2. Our data pipeline enabled large-scale feature engineering, allowing the model to learn normal user behavior and detect anomalies linked to malicious insiders. Experimental results show that adding psychometric features to technical and temporal data significantly improved accuracy, with our LSTM autoencoder achieving a precision of 0.89, recall of 0.83, F1-score of 0.86, and an AUC-ROC of 0.91. These results outperformed benchmark methods such as Isolation Forest and One-Class SVM. Overall, our findings demonstrate that fusing behavioral science with deep learning provides a more effective and scalable way to detect insider threats, offering practical value for strengthening cyber defense and guiding future research.