A Concern-Oriented Governance Approach for Securing Agentic AI

Abstract Agentic Artificial Intelligence (AgAI) embodies a substantial progression beyond predictive and generative AI, facilitating autonomous decision-making, planning, tool use, and the execution of complex, multi-step tasks with minimal human intervention across diverse domains. This increased autonomy and capability, however, also expand the potential impact and risk surface of these systems. Consequently, attacks on AgAI could lead to severe consequences, jeopardizing human lives, financial stability, and overall well-being. Thus, adherence to fundamental security properties, namely, \textit{Confidentiality}, \textit{Integrity}, and \textit{Availability} (CIA), is critical to ensuring the safe, reliable, and trustworthy deployment of AgAI. In this paper, we first identify vulnerabilities and attack surfaces in AgAI by formalizing a four-layer architecture (orchestration, perception, reasoning, and action) together with the explicit states and transitions of task execution. We subsequently map eleven primary attack classes to the three core surfaces, i.e., \textit{data}, \textit{models}, and \textit{infrastructure}, and to the specific AgAI layer where each attack is most likely to occur. Building on this mapping, we provide a detailed examination of concrete strategies and techniques for each surface. Finally, we introduce a Concern-Approach Matrix that links AgAI-specific risks to targeted countermeasures across security, AI/ML risk management, privacy/compliance, and people/organization dimensions, complemented by a lifecycle-to-layer operational checklist that translates the matrix into actionable defense mechanisms. The resulting framework offers designers, operators, and policymakers a structured, architecture-driven approach to AgAI security and governance.