MIRROR preprint reports blocking message tampering between AI agents

Researchers report in an arXiv preprint that MIRROR, a defense against message tampering between AI agents, reduced attack success rates to 0% in tests below its route-compromise threshold, at 1x LLM token cost. Tests covered three benchmarks, two frameworks, four communication topologies and a MetaGPT deployment against a production API. MIRROR sends the same message across multiple logical routes and accepts it only when a strict majority agree on its digest, a compact hash of the message. Its protection depends on an honest majority of routes: hashing alone does not authenticate messages. When routes share failure points, the largest shared-failure group matters more than the route count.