Traditional security detection methods face challenges in identifying zero-day attacks in critical infrastructures (CIs) integrated with the industrial internet of things (IIoT). These attacks exploit unknown vulnerabilities and are difficult to detect due to their connection to physical systems. The integration of legacy ICS networks with modern computing and networking technologies has significantly expanded the attack surface, making these systems more susceptible to cyber-attacks. Despite existing security measures, attackers continually find ways to breach these operating networks. Anomaly detection systems are critical in protecting these CIs from current cyber threats. This study investigates the effectiveness of unsupervised anomaly detection models in detecting operational anomalies that could lead to cyber-attacks, thereby disrupting and negatively impacting quality of life. We preprocess the data with a focus on cybersecurity and chose the SWAT dataset because it accurately represents the types of attack vectors that critical infrastructures commonly encounter. We evaluated the performance of isolation forest (IF), local outlier factor (LOF), one-class SVM (OCSVM), and Autoencoder algorithms—trained exclusively on normal data—in enhancing cybersecurity within IIoT environments. Our comprehensive analysis includes an assessment of each model’s detection capabilities. The findings highlight the VAE-LSTM model’s potential to identify cyber-attacks within seconds in a high-frequency dataset, suggesting near real-time detection capability. The final model combines the reconstruction ability of the variational autoencoder (VAE) with regularization using the Kullback–Leibler divergence, reflecting the non-Gaussian nature of industrial system data. Our model successfully detected 23 out of 26 attack scenarios in the SWAT dataset, demonstrating its effectiveness in improving the security of IIoT-based CIs.