AgentID-ZT: An Agentic Zero-Trust Framework for Adaptive Digital Identity Risk Assessment and Access Control
Abstract
: Digital identity security has become critical as modern systems rely on distributed authentication and continuous access to cloud services. Traditional mechanisms, such as static authentication and rule-based access con-trol, are increasingly ineffective against dynamic threats including credential theft, session hijacking, identity spoofing, and insider misuse. Although zero-trust architectures improve security through continuous verification, many existing approaches still lack adaptive reasoning and real-time response capabilities. This paper presents Agentic Zero-Trust (AgentID-ZT), an agentic zero-trust framework for adaptive digital identity risk assessment and access control. The framework combines dynamic identity trust graph modeling, Bayesian risk estimation, multi-agent security reasoning, and utility-driven adaptive access control. Trust relationships among users, devices, sessions, applications, and resources are modeled as a dynamic graph, while a Bayesian engine estimates identity risk from authentication, device, behavioral, and contextual signals under uncertainty. Multiple agents then collaborate to support risk detection, trust scoring, policy reasoning, and response planning. We also introduce IdentityBench-2026, a hybrid benchmark built from Los Alamos National Laboratory authentication logs and CERT insider-threat data, enriched with contextual identity signals and policy decision labels. Experiments show that AgentID-ZT outperforms all baselines, achieving 96.27% Accuracy, 95.84% F1-score, and 0.978 Area Under the Curve (AUC), with low False Acceptance Rate (FAR) 2.31% and False Rejection Rate (FRR) 3.07%. It also maintains F1-scores above 94.9% under dynamic threat scenarios and achieves an average decision latency of 41.8 ms, demonstrating practical feasibility for adaptive digital identity protection in zero-trust environments.