The integration of Internet of Things (IoT) technologies in agriculture has advanced precision farming by enabling real-time monitoring and data-driven decision-making. However, the growing reliance on interconnected sensors introduces challenges such as cybersecurity risks, sensor failures, and data irregularities that can threaten operational reliability. This study presents an IoT-based anomaly detection framework designed to enhance the security and efficiency of smart agriculture systems. The approach employs unsupervised machine learning techniques, specifically a hybrid of Principal Component Analysis (PCA) and Isolation Forest for detecting anomalies in environmental sensor data. A publicly available smart agriculture dataset containing diverse parameters like soil moisture, temperature, humidity, and light intensity was used. The model was evaluated using accuracy, precision, recall, and F1-score metrics. The results show that the PCA + Isolation Forest model achieved a high accuracy of 98.2% and a recall of 99.4%, indicating its effectiveness in detecting true anomalies while minimizing false negatives. This performance surpasses that of standalone models such as PCA, Isolation Forest, and One-Class SVM. The proposed framework is computationally efficient and well-suited for resource-constrained IoT environments commonly found in agricultural settings. By effectively identifying data irregularities, this approach enhances the security, reliability, and operational integrity of smart farming systems, making it a practical solution for supporting sustainable and secure precision agriculture.