The Attack and Defense Landscape of Agentic AI: A Comprehensive Survey

arXiv 2603.11088·medium signal

First systematic survey of AI agent security, covering the expanded attack surface when LLMs become integrated hybrid systems with tools, memory, and multi-agent pipelines. Catalogs attacks at each component — planning, memory, tools, inter-agent communication — and surveys defenses including information flow control, taint tracking, and architectural isolation (planner separated from untrusted third-party data). A synthesized reference for anyone building secure agent systems in 2026.