New Theory Proposes 'Least Autonomy' for Agentic AI Security.

A new theoretical framework, "least autonomy," is proposed to enhance security for increasingly autonomous AI systems. The authors argue that the long-standing principle of "least privilege," which limits permissions to only what's strictly necessary, is insufficient for agentic AI. This is because AI agents can dynamically combine, approve, and amplify their permissions across various workflows and system boundaries, creating new security challenges. The proposed theory formalizes this concept by defining a compositional "blast radius" that measures the structural separation between actions within an enterprise hierarchy, incorporating confidentiality, integrity, and control-context labels. It also introduces a directed "agent influence graph" to model how agents interact and influence each other, based on shared resources and communication. Finally, a "collusion predicate" is developed to detect unauthorized authorization composition, decision manipulation, and cross-domain capability integration, providing a more robust security model for complex AI deployments.