Skip to content
NewA calmer, faster ZeroHour

One malicious agent can collapse decentralized multi-agent LLM tasks

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

arXiv cs.CRby Matthew D. Laws ()1 min read
AI summary · x-ai/grok-4.7Summaries by Vani AI

An auditor attributes 98% of HiddenBench attacks, including protocol-compliant ones, with negligible trace latency.

A paper systematizes threats from malicious agents in decentralized multi-agent LLM systems that decompose and delegate tasks without a central coordinator. One malicious agent caused up to 100% task failure, raised token use by up to 7.7x and runtime by up to 3.2x, and exfiltrated up to 94% of relevant prior tasks. A cryptographic execution-trace protocol and auditor detected and attributed 98% of attacks on HiddenBench.

Unconfirmed: Attack success rates are the authors' evaluation results, not reports of real deployments.

Key facts
  • Evaluations used gpt-6.1-sol and Qwen3.8-27B-FP8.
  • Attacks target integrity, availability, and confidentiality.
  • The auditor attributes attacks that still follow the protocol.
  • No single operator can see or enforce behavior across agents.
Security assessment · AI, from the article

Impact: task failure, higher compute use, and prior-task exfiltration

Explain this AI news with Vani AIWhat was announced, what is new and why it matters, in a new chat

From the original article

1 min read at arxiv.org

Decentralized multi-agent LLM systems enable independently administered agents, each with its own operator, model, memory, and configuration, to collaborate on complex tasks without centralized coordination. Agents autonomously decompose tasks, execute subtasks and delegate remaining work to peers, enabling scalability and eliminating single points of failure.

arXiv cs.CR

By Matthew D. Laws · arxiv.org

Read the full article at arXiv cs.CR

A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.