The Agentic Trust Fabric: A Systematization of Security Architectures for Autonomous AI Agents

Autonomous agents built on large language models (LLM) interpret natural language, generate probabilistic outputs, and invoke tools as context evolves. These behaviors strain identity and access management systems built for stable and predictable principals. We systematize mechanisms across identity, authorization, runtime enforcement, and policy, and identify a maturity gap. Deployed controls are agent-agnostic, while agent-specific proposals lack production validation. We prove that stacks evaluating these dimensions independently suffer irreducible error under cross-dimensional attacks, where each dimension appears benign alone but their combination is harmful. Controlled experiments support this result. We distill design patterns, give deployment guidance with residual risks, and outline openproblems.