Source record

Published by arXiv cs.MA (T1 source). The original is at https://arxiv.org/abs/2610.02349.

Pipeline notes

The summary and note below are generated by the signal pipeline — they are Beyond Desk’s reading, not quotations from the source.

MIRROR is a communication-layer integrity primitive for LLM multi-agent systems that defends against Agent-in-the-Middle attacks by replicating canonicalized payloads across k routes and requiring a strict majority quorum on message digests. Evaluation across MMLU, HumanEval, and MBPP on four topologies and a MetaGPT deployment shows 0% attack success rate below the route-compromise threshold α<0.5 at 1x LLM token cost.

Concrete quorum-based defense for multi-agent LLM communication with measured 0% ASR at 1x token cost, compared to 44.2% benign blocking for LLM-as-a-Judge baselines.

Cited by

No citations on record.