Agentic AI security: Threat modeling + Controls for AI agents (Permissions, Tool-use Constraints, Auditability, kill-Switch Design)

Abstract

The rapid emergence of agentic artificial intelligence systems introduces a fundamental shift in cybersecurity risk, as autonomous AI agents increasingly possess persistent memory, goal-driven planning, and the ability to invoke external tools and services. Unlike traditional AI models, agentic systems act continuously within operational environments, expanding the attack surface and challenging existing security and governance frameworks. This article examines security risks specific to agentic AI through a structured threat modeling lens, focusing on vulnerabilities arising from autonomy, recursive decision-making, and multi-agent interaction. It further analyzes technical and organizational control mechanisms designed to constrain agent behavior, including permission scoping, tool-use restrictions, auditability, and safe interruption mechanisms.
By synthesizing recent research on agent architectures, trust and risk management, and ethical governance, the article argues that securing agentic AI requires a layered defense strategy that integrates architectural safeguards, runtime controls, and institutional oversight. The study contributes to the growing literature on AI security by clarifying how control surfaces and governance mechanisms can be systematically designed to ensure accountability, resilience, and safe deployment of autonomous AI agents in real-world systems.