AI Guardian and Gatekeeper: A Public Interest Architecture for Trustworthy Autonomous Agents

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

This public concept paper proposes an interoperable security and accountability architecture for autonomous AI agents. The framework combines five connected functions: an AI Guardian that supervises agent behavior; an AI Gatekeeper that verifies identity and authority before consequential interactions; a portable Agent Passport containing short-lived, cryptographically verifiable credentials; a containment system that limits damage when an agent exceeds its authority; and a forensic process that determines what occurred, assigns responsibility, and establishes conditions for reinstatement. The paper defines rogue behavior operationally as an agent attempting or performing actions outside its authorized objective, permissions, methods, recipients, resource limits, time window, or safety boundaries. The framework emphasizes least privilege, independent supervision, graduated intervention, evidence preservation, human accountability, and cross-organizational trust. The concept is offered for unrestricted public-interest development, testing, standardization, improvement, and implementation.