Tag
This paper introduces the Agent-Native Immune System (ANIS), a biologically inspired, endogenous defense architecture embedded directly within the agent's cognitive loop. It proposes a six-layer Immune Tower, a unified taxonomy of Agent Viruses and Vaccines, and the Harness Triad for continual immune learning to address runtime hijacking vulnerabilities in autonomous agents.
This paper introduces AgentWall, a runtime safety layer for local AI agents that intercepts actions before execution, enforces declarative policies, requires human approval for sensitive operations, and logs tamper-evident trails. It is open-source and works with multiple agent platforms.