Embedding an Immune System Inside AI Agents to Block Runtime Hijacks
Autonomous agents with memory, tools, and multi‑agent coordination face new attacks—memory poisoning, tool‑chain manipulation, and protocol hijacking. The paper proposes the Agent‑Native Immune System (ANIS), a six‑layer, biologically‑inspired defense embedded in the agent’s reasoning loop, with a taxonomy of agent viruses and vaccines. ANIS shifts security from static alignment to active runtime enforcement, enabling safer self‑monitoring AI.