Zero Trust Architectures Empowered by AI: A Paradigm Shift in Cloud and Edge Cybersecurity

The cloud-edge computing phenomenon and distributed digital infrastructure have exposed inherent limitations within perimeter-based security paradigms. Zero Trust Architecture (ZTA) on the "never trust, always verify" tenet has emerged as the underlying paradigm to combat advanced cyber threats. In this paper, we propose a new AI-based Zero Trust framework (AI-ZTA) that combines Transformer-based deep anomaly detection, Graph Neural Network (GNN)-based trust propagation, and Large Language Model (LLM)-assisted policy adaptation to deliver continuous, contextual, and intelligent access control in dynamic cloud-edge environments. We also incorporate Federated Learning (FL) for facilitating collaborative and privacy-preserving model training across heterogeneous edge nodes. Experimental performance on three benchmarking datasets—CICIDS2017, NSL-KDD, and an industrial cloud-edge log dataset—demonstrates that AI-ZTA identifies threats with 98.7% accuracy, a false positive rate (FPR) of merely 1.2%, and a 42% reduction in access policy violation events in comparison to traditional ZTA models. The Transformer-based behavioral engine outperforms LSTM and CNN baselines by +6.4% in precision and +8.1% in recall, and GNN-based trust modeling enhances threat context explainability by 35%. Our incorporation of LLM also enables 90% accurate auto-generation of fine-grained access policies, reducing administrative overhead by 60%. These results make AI-ZTA a scalable, resilient, and intelligent security model superior to comparable architectures and a new benchmark of Zero Trust deployment in cloud and edge computing. The proposed framework is a paradigm shift towards AI-native Industry 4.0, next-generation digital services, and critical infrastructure cybersecurity.