Skip to content
InfoResearchPreprintLLM-specific

Attacks and Defenses in Decentralized Multi-Agent Systems with Hop-by-Hop Task Decomposition and Delegation

Published
Record updated
View JSON

Summary

This research systematizes threats from malicious agents in decentralized multi-agent LLM systems, where agents decompose tasks and delegate work to peers. The authors show that a single malicious agent can cause up to 100% task failure, inflate token consumption by up to 7.7x and runtime by up to 3.2x, and exfiltrate up to 94% of relevant prior tasks from benign agents. They also present a cryptographic protocol that produces verified execution traces and an auditor that detects and attributes malicious behavior, catching 98% of attacks on HiddenBench.

Mitigation

They present a cryptographic protocol that produces an irrefutable verified execution trace of every action, and an auditor that analyzes these traces to detect and attribute malicious behavior. The protocol incurs negligible latency overhead.