Attacks and Defenses in Decentralized Multi-Agent Systems with Hop-by-Hop Task Decomposition and Delegation
4.40T1 sourcearXiv cs.MA
Source record
Published by arXiv cs.MA (T1 source). The original is at https://arxiv.org/abs/2611.25056.
Pipeline notes
The summary and note below are generated by the signal pipeline — they are Beyond Desk’s reading, not quotations from the source.
SummaryThe paper studies malicious-agent risks in decentralized LLM systems that decompose, execute, and delegate tasks without central control. It reports attacks causing task failure, higher token use and runtime, and task-history exfiltration, then proposes a cryptographic execution-trace protocol and auditor that detected and attributed 98% of tested attacks with negligible latency overhead.
Why it mattersUseful as a threat model and design checklist for operators coordinating agents across administrative boundaries, especially before enabling peer delegation in production.
Cited by
No citations on record.