Digest · I·3

Red-Teaming LLM Multi-Agent Systems via Communication Attacks

Agents in a multi-agent system coordinate through messages, and those messages may cross networks an attacker can reach. Agent-in-the-Middle (AiTM) places an LLM adversary on the channel to one victim agent, where it intercepts and rewrites messages using a reflection loop, without changing any agent. On AutoGen and CAMEL across four communication structures, it succeeds in over 40% of cases in every setting and over 70% in most.

LevelInput
ScopeMulti-agent
Type · topicAttack · Prompt Injection
Reading time3 min

Purpose

Communication is what makes an LLM multi-agent system (LLM-MAS) work, and also what makes it risky: harmful information can spread from agent to agent and amplify through the system.

Earlier work looked at the agents themselves, either turning a benign agent into a malicious one or feeding adversarial inputs to an agent. Nobody had attacked the channel between agents. That channel matters in decentralised systems, where agents run on different servers for different purposes and talk over networks that can be eavesdropped.

Approach

Agent-in-the-Middle (AiTM) places an LLM-based adversarial agent on the channel to one victim agent. It intercepts the messages meant for the victim and rewrites them.

Other agentssend normal messages
Adversarial agentintercepts, reflects, rewrites
Victim agentacts on altered messages
Whole systemrefuses service or runs injected behaviour
No agent’s profile or abilities change; only what one agent hears.

Threat model

  • Goals: denial of service, stopping the victim agent from doing its job; and targeted behaviour, making the system produce prescribed output such as injected malicious code.
  • Limited capability: the attacker controls only the messages to the victim. Other agents, their communication and external databases or tools are out of reach.
  • Limited knowledge: the attacker knows the task but not the system’s internal setup, such as its communication structure or which models the agents use.

Under those limits, and with messages that must fit each role’s format, the adversarial agent uses a reflection mechanism to write malicious instructions that fit the conversation’s context.

The paper asks three questions: is LLM-MAS vulnerable to AiTM, what makes the attack more effective, and does it work on real applications?

Claims

  • AiTM is consistently effective: attack success is above 40% in every case and above 70% in most experiments, highest on code and denial-of-service tasks.
  • It works while leaving every component of the system unchanged, including agents’ profiles and capabilities.
  • Three factors decide success, whatever the structure:
    • the position of the victim agent in the communication structure;
    • how persuasive the adversarial agent is, with more persuasive adversaries succeeding far more often;
    • the models inside both the adversarial and the victim agents.
  • On real applications it works against MetaGPT but not ChatDev.

Evaluation

The attack is tested on AutoGen and CAMEL across four communication structures: chain, tree, complete (every agent connected) and random.

Data Use
MMLU (biology and physics) Knowledge questions
HumanEval Hand-written programming problems
MBPP Programming tasks

The metric is attack success rate. Targeted behaviour counts as a success when the output contains the prescribed behaviour, such as a specific transformation for MMLU or a planted safety_check function for HumanEval and MBPP. Denial of service counts as a success when the response is a refusal like “I cannot assist with this request”.11 The paper does not say how the datasets were preprocessed, and two frameworks is a narrow base for claims about LLM-MAS in general.

Synthesis

What it adds to the taxonomy. AiTM is an input-level attack at the multi-agent scope. The poison arrives at inference as messages, the agents’ weights and configuration never change, and the damage spreads through the system. It complements attacks that plant a malicious agent: here every agent stays honest and only one channel is compromised.

Why it matters. Multi-agent systems trust their own messages. Once the channel is reachable, an attacker does not need to break any single agent’s safety training to steer the system.

Open questions:

  • Testing on more complex communication structures than the four used here.
  • Defences: authenticating or checking messages between agents is the obvious next step, and the paper leaves it open.