Security and Privacy Challenges in Agentic AI Systems

Agentic Artificial Intelligence (AI) systems can plan multi-step tasks, use external tools, retrieve information, andtake actions with limited human intervention. These capabilities create security and privacy risks beyond those found inconventional question-answering systems. An agent may encounter malicious instructions in retrieved content, misuse anauthorized tool, expose confidential context, retain unnecessary personal data, or continue a harmful action through a chain ofotherwise ordinary steps. This paper presents a risk-aware framework for analysing these challenges across the agent lifecycle. Itorganizes the main risks into instruction and prompt injection, excessive agency, insecure tool integration, data leakage, identityand access-control weaknesses, memory risks, and insufficient monitoring. The proposed approach combines least-privilegepermissions, separation of trusted instructions from untrusted data, explicit approval for high-impact actions, privacy-awarememory, output validation, and audit logging. A staged evaluation plan is described using threat scenarios and measurablesecurity and privacy indicators. This paper is a design-oriented study and does not claim experimental results or a deployedsecurity product. Its central argument is that agentic AI should be treated as a socio-technical system whose safety depends on themodel, tools, data, permissions, and surrounding operational controls