Evaluating Fine-Grained Authorisation Strategies for Securing AI Agents That Interact With Enterprise Applications

Abstract

Autonomous artificial intelligence agents equipped with large language model cognitive cores are rapidly transitioning from passive advisory interfaces into active operational actors capable of querying enterprise databases, triggering API workflows, and modifying mission-critical enterprise state. However, granting autonomous agents broad or coarse-grained access tokens introduces severe security vulnerabilities, including indirect prompt injection, confused deputy attacks, and uncontrollable privilege escalation across integrated enterprise applications. This study presents a rigorous empirical evaluation of fine-grained authorization strategies specifically architected for autonomous agent workflows. We evaluate four distinct authorization paradigms: static role-based access control, dynamic attribute-based access control, graph-based relationship access control, and a novel cryptographic delegation capability token architecture. Through extensive benchmarking across five distinct enterprise application scenarios comprising more than two million automated agent transactions, we evaluate decision latency, policy evaluation throughput, least-privilege enforcement precision, and resilience against adversarial prompt manipulation. Our findings demonstrate that while traditional attribute-based models suffer significant latency penalties under deep contextual evaluations, scoped delegation tokens combined with cryptographic boundary attestation achieve sub-ten-millisecond decision latencies while maintaining zero unauthorized state modifications across complex transactional workflows.