Agent-Native Immune System Boosts Autonomous AI Security.

As AI systems evolve from simple chatbots to complex autonomous agents with memory, tool-use, and multi-agent collaboration, the scope of potential threats has significantly expanded. Current security measures, such as perimeter defenses and training-time alignment, are often external to an agent's active reasoning and prove insufficient against runtime attacks like memory poisoning or tool manipulation. To combat this, researchers propose the Agent-Native Immune System (ANIS). ANIS is a biologically inspired, internal defense architecture integrated directly into an agent's cognitive processes. The framework introduces a six-layer "Immune Tower," including a "Barrier Immunity" layer for physical and logical isolation. It also establishes a formal taxonomy for "Agent Viruses" and "Agent Vaccines," distinguishing between superficial and robust defenses. A "Harness Triad" (Meta, Self, and Auto) provides a self-monitoring backbone for "Continual Immune Learning," allowing vaccines to adapt to new threats dynamically. The paper emphasizes that ANIS complements model alignment, which provides static value foundations, by acting as a dynamic "law enforcement" mechanism during runtime, ensuring agent security in real-time.