Generative AI-Driven Anomaly Detection for Autonomous Cloud Infrastructure Management

Cloud infrastructures generate millions of telemetry events per second, making manual monitoring impractical. Traditional anomaly detection systems rely on static thresholds or supervised ML, which limits their ability to detect novel incidents. This study presents a generative AI framework that uses variational autoencoders and diffusion-based sequence models to learn normal cloud behavior and flag emerging anomalies. The model processes log streams, metrics, configuration changes and request traces from distributed microservices. Experiments were conducted on OpenStack clusters and AWS EC2 testbeds under both normal and fault-injected conditions. Results show a 29 percent improvement in anomaly detection accuracy and a 33 percent reduction in false alarms compared with LSTM- and PCA-based baselines. The system also predicts incident severity and suggests remediation actions through a reinforcement learning component trained on historical operational data. Deployment studies confirm that the model scales efficiently across multi-region cloud clusters and maintains sub-100 ms detection latency. The findings demonstrate that generative AI can significantly improve cloud observability and autonomous IT operations. Future work will focus on integrating cross-platform observability standards such as OpenTelemetry.