Developer Builds Experimental Cross-Agent AI Threat Detection System, Exposes Key Flaws

A developer created ImmuneMesh, a research prototype designed to detect self-replicating prompt attacks spreading across multi-agent AI systems, a vulnerability current guardrail tools largely ignore. The system works as LangGraph middleware, using a shared 'antibody store' so that once one agent identifies a threat, all other agents in the mesh are immediately alerted. Testing against eight benign task categories revealed a significant false-positive problem, with legitimate tasks like proofreading and instruction confirmation being incorrectly flagged as threats. To address this, the developer introduced per-agent adaptive baselines, though building the fix exposed a bootstrap deadlock where agents with naturally high mirroring scores could not accumulate enough history to set their own normal threshold. The creator acknowledges the prototype is not adversarially hardened and that baseline poisoning remains an unsolved research challenge.