From Data Leakage to Intent Leakage: A Hierarchical Risk Taxonomy for Agentic AI Systems

As AI systems evolve from passive inference engines to autonomous agents, the risk landscape shifts from dataexposure to decision exposure. This paper proposes a five-level hierarchical leakage taxonomy for agentic AI thatextends established data and information leakage categories to include cognitive leakage, intent leakage, andconjunctive leakage. We further introduce Decision Reconstruction as an adversarial primitive — the systematic recovery of decisionlogic from observable agent behavior — extending side-channel analysis principles to the semantic level of agenticdecision-making. The taxonomy and gap analysis demonstrate that current governance frameworks lack explicitdefinitions and assessment criteria for these risk categories, despite increasing regulatory obligations. This work is a conceptual and governance-oriented preprint. It defines a risk taxonomy and governance gap analysiswithout prescribing implementation mechanisms or technical controls.