AgentAntibody: Adaptive Immune System for LLM Agents Against Prompt Injection

A new research paper on arXiv (2608.04053) introduces AgentAntibody, a defense mechanism for large language model (LLM) agents against prompt injection attacks. The system is inspired by adaptive immunity, equipping agents with a self-evolving immune system that learns from each encounter. It maintains a persistent library of antibodies representing the user's security boundary, which recognizes threats and mounts immune responses. The approach addresses the issue of underspecified user requests, where injections can exploit ambiguity. The paper is a cross-announcement, indicating it has been submitted to multiple venues. The work is relevant to the growing field of AI security, particularly for autonomous agents.