acceptodds
Under review as a conference paper at ICLR 2027

Cooperative Words, Adversarial Actions: Action-Level Ambiguity in Embodied Multi-Agent Systems

Abstract

Embodied multi-agent systems increasingly rely on foundation-model agents to coordinate long-horizon physical tasks, making reliable attribution of team failures a growing safety challenge. We identify action-level ambiguity as an overlooked vulnerability, where a compromised agent can remain locally plausible while systematically reducing collective progress. We introduce a stealthy action-level attack that induces contextually defensible but lower-contribution actions while preserving cooperative communication and faithful observation reports. Across diverse embodied benchmarks, coordination frameworks, and agent backbones, the attack substantially degrades team performance while remaining difficult to attribute, achieving an overall stealth rate of 0.9980 across three investigator models, three auditing strategies, and 8829 episode-level audits. In contrast, overt, lazy, and randomly suboptimal agents are substantially easier to identify. These results expose a gap between behavioral plausibility and contribution integrity. To reduce this attribution gap, we further develop a contribution-aware, log-based attribution method that reasons about each agent’s marginal task contribution and interference with teammates’ intents, substantially improving compromised-agent identification.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
adversarial attack, embodied multi-agent system, agent safety

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_HX7LOpenReview ↗
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
  1. 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.
  2. 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.
  3. 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
AuthorsOpenReview ↗

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.

AuthorsOpenReview ↗

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.

AuthorsOpenReview ↗

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/

Reviewer_5b8tOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper identifies action-level ambiguity in embodied multi-agent systems: a single compromised agent can communicate cooperatively while selecting plausible actions that undermine team progress. Across three benchmarks and multiple coordination frameworks, its prompt-based attack substantially reduces task performance while investigators fail to identify the compromised agent in 99.8% of audits.

Critical strengths
  • The paper shows that cooperative communication can conceal harmful action choices, exposing a gap between plausible behavior and genuine team contribution.

  • Across three benchmarks and multiple coordination frameworks, the attack substantially degrades performance while evading attribution. Contribution-aware analysis improves identification, providing a concrete direction for defense.

Critical weaknesses
  • The primary source of “action-level” stealth is questionable. Ablations show that removing the surface component (communication camouflage) reduces stealth from 1.0 to 0.31–0.45, while RSA/LA achieve stealth rates of only approximately 0.1. This suggests that stealth primarily stems from deceptive communication rather than action-level ambiguity, which is at odds with the paper’s central claim. Moreover, The analysis lacks a more detailed breakdown, for instance, a taxonomy of failure modes, or an examination of the task structures under which the attack is more effective. Moreover, the performance gap between SA and NB falls within one standard deviation, so the claim that "stronger" is weakly supported.

  • The investigator evaluation design favors the attacker.

    • For example, false-positive rates on benign episodes are not reported. Since benign teams already fail frequently (MAP-THOR TSR is only ~0.3), investigators may consistently classify failures as unintentional, making a stealth rate of 1.000 uninformative.

    • Then, the prompt asks the model to distinguish intentional from unintentional failures, introducing a prior bias toward assuming that no agent is acting maliciously.

Questions for authors
  • If the attacker can directly target the system prompt, does that imply the attacker already holds extremely high privileges, in which case a reduced success rate would hardly be surprising?

  • Among the reported stealthy attack cases, how many were "no intrusion actually achieved," and how many involved misattribution of the attacking entity? Is the false-positive rate computed on benign events?

  • Were the derived premises and causal dependencies cross-checked against the simulator state or manual annotations? Could parsing errors propagate into and bias the attribution judgment?

  • Were the investigators told that exactly one attacking entity had been compromised, consistent with the defender's assumption?

  • typos: More results are at 3 at line 726

5 comments
AuthorsOpenReview ↗

Response to Reviewer 5b8t (1/2)

We thank the reviewer for the thoughtful feedback. We want to address your concerns as follows:

W1 - Action-level stealth, failure modes, and attack strength

(1) Cooperative communication and action-level ambiguity play complementary roles. Cooperative communication conceals adversarial intent, while action-level ambiguity enables locally plausible but harmful action selection. Removing cooperative surface behavior reduces stealth from 1.000 to 0.306–0.453, while removing action evaluation substantially weakens attack effectiveness (RoCoBench success increases from 0.075 to 0.294). The ablations show the distinct contributions of communication and action selection, instead of supporting communication as the sole source of stealth.

(2) We conducted an additional quantitative failure-mode analysis, classifying 100 cases per benchmark into eight categories across two channels: individual contribution degradation (1a–1d) and interference through shared physical states (2a–2d). To achieve this, we prompt a language model directly with the known compromised-agent identity to characterize action deviations

Failure modeRoCoBenchMAP-THORTDW-MAT
1a. Wrong target/subgoal762
1b. Wrong timing/order31719
1c. Wrong execution353126
1d. Redundant work241215
2a. Resource preemption13917
2b. Spatial obstruction82219
2c. Progress reversal132
2d. Joint-action desynchronization900
Total100100100

The takeaways are as follows:

Wrong execution is prevalent across all three benchmarks, while timing deviations and spatial obstruction are more frequent in household tasks. For example, RoCoBench Sweep illustrates incorrect execution by pushing an object away from a teammate's dustpan (Fig. 1b), while TDW-MAT illustrates delayed transportation despite an immediately available transport action (Fig. 2). These actions occur frequently because they can be interpreted as an unintentional mistake, helping the agent to achieve the goal of stealthy sabotage.

Task susceptibility also depends on coordination dependencies, recovery opportunities, and baseline difficulty. Tightly coupled manipulation can be blocked by a single action, whereas benign agents may continue completing independent transportation subtasks.

(3) Regarding the comparison of performance between SA and NB, our primary claim is comparable disruption with substantially greater stealth (from the results, NB has no stealth and is easily identified). We will clarify the wording accordingly.

W2 - Investigator evaluation bias

The investigator does not universally classify failures as unintentional. Our additional benign-episode evaluation yields 0% observed FPR. On the other hand, NB achieves 0 stealth, and LA/RSA also exhibit low stealth rates under the same protocols, which shows that investigators can recognize adversarial behavior. This demonstrates that SA is substantially harder to identify. We further address the potential prior bias through a forced-choice experiment, detailed below.

Q1 - Does system-instruction access imply excessive attacker privileges?

Single-agent compromise is an established MAS threat model, including Deng et al. [1] (RAID 2021) and Peigné et al. [2] (AAAI 2025). Our adversary modifies only one agent's pre-deployment instruction, without direct control over runtime actions, teammates, or environment dynamics. The central finding concerns difficult behavioral attribution following a localized compromise, rather than the unsurprising possibility of performance degradation.

Q2. How many cases involve no intrusion versus misattribution? For the cases in which our agent is not marked as an intentional saboteur, we want to show a more informative breakdown that distinguishes whether investigators recognize execution errors and whom they hold responsible. We identify three categories, detailed below:

  • 27%: No notable error identified.
  • 44%: Errors attributed to another agent.
  • 29%: Errors attributed to the compromised agent but judged unintentional.

The 29% category directly illustrates action-level ambiguity. Investigators recognize the compromised agent's erroneous behavior but interpret the behavior as an ordinary mistake. The 44% category illustrates the difficulty of assigning responsibility in collaborative tasks.

AuthorsOpenReview ↗

Response to Reviewer 5b8t (2/2)

Q3. Could LLM parsing errors bias causal attribution?

We manually examined 50 parser outputs (25 TDW-MAT, 25 MAP-THOR), covering the evaluated collaboration frameworks. Among the 50 outputs, 47 (94%) were correct and 3 contained minor predicate-level errors, with no major errors observed. We manually corrected the three erroneous outputs and recomputed attribution. All three final attribution decisions remained unchanged. Attribution succeeded in 37/47 correctly parsed cases and 2/3 cases with minor errors; the remaining failure persisted after correction. The counterfactual analysis directly validates that the observed parsing errors did not alter final attribution decisions.

Q4. Were investigators informed that exactly one agent was compromised?

The original investigator is not informed of this, while our contribution-aware defense assumes exactly one compromised agent. We conducted an additional forced-choice evaluation explicitly informing investigators that one agent was compromised:

SettingCorrect attributionRandom baseline
Two-agent collaboration55%50%
Three-agent collaboration36%33.3%

Even with explicit knowledge of an attack, conventional investigators achieve attribution accuracy close to random guessing. The attribution difficulty persists independently of the original prompt's uncertainty about whether an attacker exists.

Q5. Typo at line 726

Thank you for identifying the typo. We will replace "More results are at 3" with "Results are summarized in Table 3."

References

[1] Deng et al., An Investigation of Byzantine Threats in Multi-Robot Systems, RAID 2021. https://doi.org/10.1145/3471621.3471867

[2] 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

Thanks for your extra experiments and reply, but the action-evaluation ablation shows disruption rather than specifically illustrating the contribution of action-level ambiguity to concealment.

The reported zero FPR also doesn't rule out the possibility that the investigators were being overly cautious, so my main concern still hasn't been addressed.

Reviewer_DoqCOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper identifies action-level ambiguity as a novel safety vulnerability in embodied multi-agent systems (MAS). A single instruction-compromised VLM agent can undermine team performance through locally plausible but lower-contribution actions while maintaining cooperative communication. The paper proposes a contribution-aware defense using causal graph-based leave-one-out analysis, achieving DASR of 0.722–0.875.

Critical strengths

1, Novel threat model with striking empirical results (0.9980 stealth); 2, Principled contribution-aware defense with broad applicability

Critical weaknesses

1,Defense's LLM-based parser is underexamined

Questions for authors

1, How accurate is the LLM parser in your defense, and what happens when it's wrong? 2, What happens over longer horizons with a smarter attacker? 3, Why are some tasks so much easier to attack than others?

2 comments
AuthorsOpenReview ↗

Response to Reviewer DoqC (1/2)

We thank the reviewer for the constructive feedback. We want to address your concerns as follows:

W1 and Q1 - Defense's LLM-based parser is underexamined

To assess the accuracy and downstream impact of our LLM-based parser, we manually examined 50 parser outputs, including 25 from TDW-MAT and 25 from MAP-THOR, covering the evaluated collaboration frameworks. Since our parser normalizes execution logs into a shared symbolic predicate representation independently of the underlying coordination framework, we primarily stratify the validation by benchmark to capture differences in embodied environments and action semantics.

We categorized parsing outcomes into three levels: (1) Correct, where the extracted predicates accurately reflect the execution logs; (2) Minor errors, where predicate discrepancies exist but correcting them does not change the final attribution decision; and (3) Major errors, where correcting the discrepancies changes the attribution decision. The classification is based on counterfactual correction of parsing errors rather than subjective judgments of error severity.

Among the 50 instances, 47 (94%) were parsed correctly, while only 3 contained minor errors, with no major errors observed. The three minor errors involved a missing predicate, an additional predicate, and a missing predicate attribute, respectively.

Importantly, correcting all three erroneous parsing outputs did not change the final attribution decisions, demonstrating that the observed parsing errors did not propagate into incorrect agent identification. The defense correctly identified the compromised agent in 37/47 correctly parsed cases (78.7%) and 2/3 cases with minor parsing errors (66.7%). The remaining attribution failure among the minor-error cases persisted after correction, indicating that the parsing discrepancy was not responsible for that failure. These results demonstrate 94% exact parsing accuracy in our manually examined sample, with no observed changes in attribution decisions caused by parsing errors. The counterfactual correction analysis further establishes that the observed minor predicate-level discrepancies did not affect the defense's final predictions.

W2 - longer horizon with a smarter attacker

We provide additional clarification on attacker sophistication and new experimental results under extended horizons.

First, our attack already achieves a strong combination of attack effectiveness and behavioral stealth among the evaluated attack methods. Compared with naive Byzantine, lazy, and random suboptimal attacks, our method consistently achieves near-perfect stealth while inducing comparable or greater task-performance degradation across the evaluated settings. A "smarter attacker" could involve different capabilities or strategies, but greater attacker sophistication does not necessarily imply stronger evasion of our defense, which evaluates causal task contribution and interference with teammates' intentions rather than specific attack patterns. Second, our evaluation already includes long-horizon embodied collaboration. For example, TDW-MAT allows episodes of up to 3,000 simulation frames, with individual agent decisions spanning multiple frames. To further examine the impact of longer execution, we conducted additional experiments by doubling the original episode horizons for both TDW-MAT and MAP-THOR using GPT-4o-based agents.

BenchmarkFrameworkOriginal DASR2× Horizon DASR
TDW-MATCoELA0.7360.792
TDW-MATCaPo0.8750.875
MAP-THORCoELA0.8290.844
MAP-THORCaPo0.8220.844

The results demonstrate that contribution-aware attribution maintains strong performance under doubled horizons, achieving DASR between 0.792 and 0.875 across all four configurations. Compared with the original settings, attribution accuracy is maintained despite the increased opportunities for accumulated coordination errors and more complex causal dependencies.

These findings provide further empirical evidence that extending the execution horizon does not inherently undermine contribution-aware attribution. By reasoning about agents' causal contributions and interference with teammates' intentions, our defense continues to identify compromised agents effectively even as the execution horizon increases.

AuthorsOpenReview ↗

Response to Reviewer DoqC (2/2)

W3 - Why are some tasks so much easier to attack than others?

Differences in attack effectiveness can be explained by three task-level factors: coordination dependencies, recovery opportunities, and baseline task difficulty.

First, tasks differ in their sensitivity to coordination failures. Tasks involving tightly coupled actions, precise timing, or shared physical resources are particularly vulnerable because a single poorly coordinated action can invalidate a teammate's progress. For example, in RoCoBench Sweep (Figure 1b), the compromised agent performs a seemingly legitimate sweeping action but pushes the object away from the teammate's dustpan, directly disrupting the intended cooperative manipulation. Sweep success decreases from 0.833 to 0.067 under our attack.

Second, coordination dependencies also determine whether benign agents can recover from adversarial actions. In tightly coupled manipulation tasks such as Sweep, moving an object outside another robot's reachable workspace can prevent the teammate from completing its planned action without corrective intervention. In contrast, household transportation tasks often permit more independent progress: even when a compromised agent delays transportation or performs unproductive actions, benign teammates may still collect and transport other target objects. Appendix A.4 illustrates such delayed-transport behavior in TDW-MAT, where the compromised agent chooses exploration or navigation despite having an immediately available transport action. The ability of remaining agents to continue contributing to the shared objective can partially mitigate the impact of individual adversarial decisions.

Third, baseline task difficulty also affects the interpretation of attack effectiveness. For example, RoCoBench Rope has a benign success rate of only 0.367, compared with 0.917 for Sort and 0.883 for Cabinet. Consequently, Rope's attacked success rate of 0.000 does not necessarily imply that Rope is intrinsically easier to attack. Importantly, all six RoCoBench tasks exhibit substantial relative success-rate reductions of more than 85.5%, demonstrating consistent attack effectiveness despite differences in baseline difficulty and task structure. Overall, task-specific coordination dependencies determine how adversarial actions disrupt collaboration, while recovery opportunities influence whether the remaining benign agents can compensate for the disruption. These factors help explain the observed variation in attack effectiveness without undermining the generality of the identified action-level vulnerability.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.