When Agents Trade: A Comprehensive Taxonomy of the AI Agent x DeFi Attack Surface

Abstract

First systematic security study of the AI Agent x DeFi intersection. Contributions: (1) Five-layer attack surface model (user input, LLM reasoning, MCP tool layer, smart contracts, blockchain state); (2) 8-vector attack taxonomy (V1-V8) with 3 unique properties (Interpretation Gap, Automation Surface, Memory Persistence); (3) Empirical validation through audit of 5 protocols (108 contracts), finding 13 vulnerabilities across 7/8 vectors; (4) Three-level defense framework (protocol, agent, infrastructure). Includes Foundry-based attack sandbox.