Digest · I·3

ACIArena: Toward Unified Evaluation for Agent Cascading Injection

In a multi-agent system, one compromised input, agent or message can cascade through the rest. ACIArena is a benchmark of 1,356 test cases that crosses three attack surfaces with three goals (hijacking, disruption, exfiltration) on six real multi-agent frameworks. It finds that topology alone does not predict robustness, that critic roles help, and that existing defences trade security for utility. It also proposes ACI-Sentinel, which trims each agent's context to what the task needs.

LevelInput
ScopeMulti-agent
Type · topicEvaluation · Prompt Injection
Reading time4 min

Purpose

In a multi-agent system (MAS), agent cascading injection (ACI) means compromising one point of the system and letting the damage spread to the rest. Research on it had three gaps:

  • Incomplete threat scenarios. Studies looked only at agent profiles or only at messages as the attack vector, or only at one goal such as system compromise or privacy leakage.
  • Simplified evaluation settings. Many used toy MAS implementations that differ a lot from real systems, so findings did not transfer.
  • Hard-to-extend code. Existing codebases were not modular enough to add new systems, attacks or defences.

MASLab already offered a unified MAS framework with standard inputs and protocols, but it lacked an evaluation environment and varied attack strategies. ACIArena fills that gap with one framework for benchmarking how robust different MAS are under different threat models.

Approach

ACIArena builds on MASLab and evaluates each system across a grid of threats.

Attack surfaceexternal input · agent profile · inter-agent message
Objectivehijacking · disruption · exfiltration
Task domainmath · code · science · medical
1,356 test casesrun on six MAS
The benchmark crosses where the attack enters with what it is trying to achieve, over benign tasks from four domains.
  • Attack surfaces: external inputs, a malicious agent (its profile), and poisoned messages between agents.
  • Objectives: instruction hijacking, task disruption, and exfiltration of sensitive data.
  • Benign tasks: math (GSM8K, MATH500), code (HumanEval, MBPP), science (GPQA) and medical (MedMCQA), filtered by an LLM judge for complexity, decomposability and low ambiguity.
  • Attack prompts are generated without gradients. Starting from a hand-written objective, mutation operators produce variants, which are run against several MAS. The variant that scores best on stealth (looking benign) and harmfulness (achieving the objective) is kept.

Claims

  • Comprehensive: 1,356 test cases over six widely used MAS implementations and every combination of attack surface and objective.
  • Standardised: one interface for implementing systems, attacks and defences, so results are comparable.
  • Extensible: a modular design that makes adding new systems or defences quick.

The main findings:

  • Judging robustness from topology alone is not enough. The same communication graph can give very different results depending on how roles are assigned.
  • Critic roles help. Systems with a critic, as in AgentVerse and CAMEL, are generally more secure; complex systems without one can be even more fragile than simple ones.
  • Code generation is highly vulnerable. Many systems reach 90–100% attack success under hijacking or disruption.
  • A stronger model does not mean a safer system. Stronger models showed higher success rates in the disruption suite.
  • Spreading roles spreads risk. Every agent becomes a potential single point of failure.
  • Utility and security trade off. Defences that block attacks often break the task too.11 This trade-off is the thread through the whole paper: almost every defence that helps security costs utility somewhere.

Evaluation

The six systems are MetaGPT, AutoGen, CAMEL, Self-Consistency, LLM Debate and AgentVerse, run with GPT-4o, GPT-4o-mini and Qwen2.5-7B-Instruct.

Metric What it measures
Benign utility (BU) Task success with no attack (pass@1)
Attack success rate (ASR) Share of tasks where the attack objective is met, by string match or an LLM judge
Utility under attack (UA) Task success while under attack, comparable with BU
Propagation Vulnerability Index (PVI) How far an attack spreads: each agent’s ASR weighted by its distance to the final answer

Five defences serve as baselines: a DeBERTa prompt-injection detector, delimiters (spotlighting), sandwich prompting, the AGrail guardrail and G-Safeguard. The detector and delimiters underperform. Sandwich prompting helps in places but can make exfiltration worse. AGrail improves security at a near-total loss of utility, and G-Safeguard costs utility for limited gains.

The paper also proposes ACI-Sentinel, which after each step trims an agent’s context down to what the task needs (“contextual least privilege”) instead of trying to spot malicious messages. It beats the baselines, for example cutting ASR by 53.33% for AutoGen under exfiltration and 79.17% for MetaGPT under hijacking, but still fails on some suites and costs utility on complex tasks.

Synthesis

What it adds to the taxonomy. ACIArena is an input-level evaluation at the multi-agent scope: every attack arrives at inference, through inputs, profiles or messages, and no weights change. It is the benchmark the other multi-agent prompt-injection papers can be measured against.

Why it matters. Current multi-agent systems remain highly susceptible to cascading injection, and defences built in simplified settings often fail to transfer. Even the proposed defence is far from a complete answer.

Limitations and open questions:

  • Query costs kept the experiments small, so newer models such as GPT-5 and Gemini 3 are not covered.
  • The work improves robustness only through external defences, not through how the system itself is designed. The authors argue that robustness should be built into MAS design, through deliberate roles and controlled interaction, rather than added afterwards.