- Rating
- 1(−1)
- Confidence
- 4(+1)
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 1(−1)
Read review
Paper summary
This paper investigates a covert attack method in which an agent consistently appears cooperative during verbal communication while, in reality, performing actions that progressively diminish the efficiency of the team's task execution. And the feasibility of this approach is further validated through experiments.
Critical strengths
Extending the concept of adversarial attacks from traditional multi-agent systems to embodied multi-agent systems is a sound approach, and the problem itself is worthy of investigation. The experimental validation is also quite thorough.
Critical weaknesses
- The method is lack of novelty. It simply migrated the prompt attack from LLM to the embodied multi-agent system, using the method of giving the LLM non-stealthy erroneous instructions to cause the embodied multi-agent system to malfunction.
- The threat model is overly stringent, making it highly difficult for an attacker to directly target system instructions. Any malicious instructions introduced by an attacker would lack any form of stealth.
- The problem is meaningful, but the method is meaningless. In traditional multi-agent systems, the core challenge lies in determining the specific actions required to execute an adversarial attack. I believe the same applies to embodied multi-agent systems: rather than relying solely on a single prompt, one should employ optimization techniques to derive effective adversarial actions. It would be a significant improvement if the authors explored using gradient-based methods to train malicious prompts while preserving their original semantic meaning. This would offer much greater value than the current approach.
Questions for authors
1.What is the novelty of the proposed attack compared to existing prompt-based attacks on LLMs? 2.Why do the authors rely on a hand-crafted prompt rather than optimizing adversarial actions or prompts through methods such as gradient-based optimization while preserving benign semantics?
13 comments
Response to Reviewer HX7L (2/4)
Weakness 2 — Threat Model and Behavioral Stealth
We respectfully disagree with the reviewer's characterization of our threat model as unrealistic and the assumption that malicious instructions necessarily eliminate attack stealth. The criticism conflates control over one agent's instruction context with unrestricted control over the entire multi-agent system, and instruction-level visibility with behavioral detectability.
First, single-agent compromise is an established and practically motivated threat model in multi-agent security research. Deng et al. [3] (RAID 2021) investigate Byzantine threats originating from untrusted robots in multi-robot systems. Peigné et al. [4] (AAAI 2025) explicitly consider an attacker compromising one agent to influence an entire LLM-based multi-agent system, while Arif et al. [5] (ACL 2026) study adversarial instructions introduced through a single compromised remote agent. These studies show single-agent compromise as a recognized security threat without requiring system-wide control. Real-world deployments further reinforce its practical relevance. For example, Changi Airport coordinates heterogeneous robots from multiple manufacturers [6], while the Amazon Q Developer supply-chain incident (CVE-2025-8217) demonstrates a concrete pathway for introducing malicious instructions through compromised software components [7]. Consistent with these settings, our adversary controls only one agent's pre-deployment instructions, without direct access to runtime actions, other agents, observations, communication channels, or environment dynamics.
Second, the trust boundaries inherent in independently developed or configured agents introduce a distinct behavioral attribution challenge. In heterogeneous multi-agent deployments, external investigators may observe inter-agent communication and execution trajectories without possessing privileged access to every participant's internal instructions. Under such conditions, identifying a compromised participant requires distinguishing adversarial actions from legitimate planning errors and ordinary coordination failures. The visibility of malicious instructions under privileged inspection provides no evidence that malicious behavior can be identified from externally observable execution records. Our work investigates precisely the gap between instruction-level compromise and behavior-level attribution, where a compromised agent can preserve cooperative communication while selecting locally plausible actions that undermine collective progress.
Third, extensive experiments demonstrate that behavioral stealth is achieved through deliberate attack design. Across three investigator models, three auditing strategies, and 8,829 episode-level audits, our attack achieves an overall stealth rate of 0.9980. Removing the cooperative surface-behavior component reduces stealth rates from 1.000 to 0.306–0.453, establishing the importance of communication camouflage. Meanwhile, removing context-dependent action evaluation substantially weakens attack effectiveness, confirming the complementary role of structured action selection in disrupting collective performance.
Response to Reviewer HX7L (3/4)
Weakness 3 and Question 2— Methodological Significance and Gradient-Based Optimization
We strongly disagree with the reviewer's assertion that our method is "meaningless" without gradient-based optimization. The criticism overlooks the technical infeasibility of direct gradient-based optimization against closed-source VLMs and the importance of developing effective attacks under realistic adversarial capabilities.
First, direct gradient-based optimization is infeasible under closed-source model access. Our experiments employ GPT-4o, whose model parameters and gradients are inaccessible to external attackers. Standard gradient-based prompt optimization requires white-box access to the target model, an exceptionally strong assumption for attacks against commercially deployed VLM-based agents. Even with model gradients, optimizing actual embodied task performance remains challenging because the attack objective depends on discrete executable actions, long-horizon environment transitions, and interactions among autonomous agents. Using open-source surrogate models would introduce additional assumptions about cross-model transferability, while estimating attack effectiveness through repeated environment rollouts requires extensive access to the target system. Such optimization requirements substantially restrict real-world attack feasibility and exceed the capabilities of an adversary who compromises a single agent's pre-deployment instructions.
Second, our method directly addresses the challenge of identifying effective adversarial actions through online, context-dependent action evaluation. Our structured instruction guides the compromised agent to identify actions that best advance the shared objective and dynamically select locally plausible alternatives with lower contributions. The agent continuously adapts its adversarial action choices to evolving observations, memory, and teammates' behavior, enabling effective sabotage without external optimization or task-specific action scripts. Our ablation results establish the importance of action evaluation: removing the corresponding component increases RoCoBench success from 0.075 to 0.294 and TDW-MAT CoELA transportation rate from 0.107 to 0.287.
Third, the effectiveness of a training-free, task-agnostic attack reveals the severity and accessibility of the identified vulnerability. A single fixed instruction intervention substantially degrades collective performance across three embodied benchmarks and multiple coordination frameworks, achieving an overall behavioral stealth rate of 0.9980 without gradient access, adversarial training, or task-specific optimization. The ability to induce effective and difficult-to-attribute sabotage under minimal attacker capabilities constitutes a significant security finding: exploiting action-level ambiguity does not require privileged model access, extensive computational resources, or repeated interactions with the target environment.
Our method demonstrates that effective adversarial actions can be generated adaptively by the compromised agent itself under realistic black-box constraints. The reviewer's proposed gradient-based optimization requires substantially stronger access assumptions and introduces practical barriers absent from our attack, while the demonstrated effectiveness of a single training-free intervention establishes the vulnerability's practical significance.
[1] Vaswani et al., Attention Is All You Need, NeurIPS 2017. https://proceedings.neurips.cc/paper/2017/hash/3f5ee243547dee91fbd053c1c4a845aa-Abstract.html
[2] Dosovitskiy et al., An Image Is Worth 16x16 Words: Transformers for Image Recognition at Scale, ICLR 2021. https://openreview.net/forum?id=YicbFdNTTy [3] Deng et al., An Investigation of Byzantine Threats in Multi-Robot Systems, RAID 2021. https://doi.org/10.1145/3471621.3471867
[4] Peigné et al., Multi-Agent Security Tax: Trading Off Security and Collaboration Capabilities in Multi-Agent Systems, AAAI 2025. https://doi.org/10.1609/aaai.v39i26.34970
[5] Arif et al., Conjunctive Prompt Attacks in Multi-Agent LLM Systems, ACL 2026. https://aclanthology.org/2026.acl-long.1577/
[6] Changi Airport Group, Meet Changi Airport's High-Tech Cleaning Robot Conductor, Changi Journeys, 2025. https://www.changiairport.com/en/corporate/our-media-hub/publications/changi-journeys/2025/robotic-middleware-framework.html
[7] Amazon Web Services, Amazon Q Developer Extension Security Update (AWS-2025-015), AWS Security Bulletin, 2025. https://aws.amazon.com/security/security-bulletins/AWS-2025-015/
Response to Reviewer HX7L (1/4)
Thank you for your detailed comments. We want to address your concerns and clarify misunderstandings:
Weakness 1 & Question 1— Novelty and Non-trivial Technical Contributions
We respectfully disagree with the reviewer's characterization that our work "simply migrates prompt attacks from LLMs to embodied multi-agent systems." This characterization conflates the attack entry point (instruction modification) with the underlying adversarial mechanism (context-dependent action-level sabotage), and overlooks the fundamental differences introduced by embodied multi-agent coordination.
First, our attack is fundamentally different from simply instructing an LLM to produce erroneous outputs. In conventional prompt attacks, malicious behavior is often directly reflected in the generated content. In contrast, our attack exploits a unique property of embodied multi-agent collaboration: an action can be individually valid, contextually plausible, and consistent with cooperative communication, yet adversarial in its consequences for collective task progress. Our structured attack explicitly induces the agent to evaluate actions that advance the shared objective and then select plausible lower-contribution alternatives based on evolving observations and teammates' behavior. This context-dependent action selection is non-trivial and is supported by our ablation results: removing action evaluation increases RoCoBench success from 0.075 to 0.294, substantially weakening the attack.
Second, our contribution extends beyond attack construction to uncovering and addressing a previously underexplored attribution problem. We demonstrate that conventional behavioral investigators struggle to identify compromised agents despite observing complete execution trajectories, achieving an overall attack stealth rate of 0.9980. More importantly, we develop a contribution-aware attribution method that explicitly reasons about causal task contribution and interference with teammates' intentions, achieving an average attribution accuracy of 79.8% across twelve configurations.
Third, transferring an existing paradigm to a new domain does not inherently imply a lack of novelty. An example is the transition from Transformers in NLP [1] (Vaswani et al., NeurIPS 2017) to Vision Transformers [2] (Dosovitskiy et al., ICLR 2021). Although ViT builds upon the Transformer architecture, its significance lies in demonstrating how the paradigm can be effectively adapted to vision, addressing domain-specific challenges and revealing new capabilities. Similarly, the novelty of our work should be evaluated based on the new security mechanisms and challenges arising in embodied multi-agent systems, instead of being dismissed solely because the initial attack intervention uses natural-language instructions.
In summary, reducing our work to a straightforward prompt attack overlooks its central technical contributions: identifying action-level ambiguity as a distinct security vulnerability, developing a structured mechanism to exploit it, and introducing a corresponding causal attribution method. These contributions cannot be obtained by simply applying existing LLM prompt attacks to embodied agents without investigating the unique action dependencies and coordination dynamics of the new setting.