acceptodds
Under review as a conference paper at ICLR 2027

Which Structures Amplify Attacks? Uncovering Collaboration-Induced Vulnerability in LLM Multi-Agent Systems

Abstract

Collaboration enables LLM-based multi-agent systems (MAS) to solve complex tasks, but can also amplify adversarial influence. We observe that direct collaboration among malicious agents strengthens existing attacks, with gains varying substantially across communication topologies. This raises a question: which collaboration structures most effectively amplify attacks? Our analysis associates higher attack success with denser internal connections that support attacker coordination and outward connections that enable propagation to normal agents. These findings motivate Collaboration-Induced Vulnerability (CIV), a perspective that examines how local collaboration structures support attack amplification. To assess CIV, we propose CIV-Amplify, a graph-based red-teaming framework that identifies vulnerable collaborative subgraphs using agent attributes, internal coordination, and outward connections. It then adapts existing attacks on multi-agent systems into coordinated behaviors tailored to the selected structure. Our work highlights structural weaknesses in multi-agent systems and supports security evaluation against collaborative threats. Our code is anonymously released at https://anonymous.4open.science/r/CIV-Amplify-6F15.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM, Multi-agent Systems (MAS), Security, Collaboration

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_GLecOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper shows that LLM multi-agent systems are more vulnerable when compromised agents can coordinate, particularly when those agents are densely connected to each other and to normal agents. They then propose CIV-Amplify, which picks a vulnerable group of attackers using a learned predictor and has an LLM planner assign them coordinated roles. Across three attack types and four topologies, this raises average ASR from about 26.58% to 36.27%.

Critical strengths
  1. The distinction between attacker-to-attacker and attacker-to-normal connectivity is a useful lens on collaborative attacks in MAS. Beyond the density-group comparisons, Fig 5 varies one density while holding the other fixed, which gives some controlled evidence that both matters. The evaluation is broad, covering three attack types, four topologies, and several backbones. The average gain in Table 1 is 9.69 points, and it is largest for memory poisoning, which suggests the effect is not limited to a single setting.

  2. Fig 6 shows that CIV-Amplify retains an advantage over the original attacks under all three evaluated safeguards. This makes the evaluation more informative than testing only undefended systems.

Critical weaknesses
  1. The predictor’s effectiveness for subgraph selection is not sufficiently established. Table 2 reports aggregate metrics but does not directly evaluate ranking among alternative attacker sets within a fixed graph and budget. The predictor is also trained on basic-collaboration ASR, while selection precedes planner optimization, which may affect candidates differently. Fig 7 provides some downstream support, but within-graph selection results after planning and a trained structure-only baseline would better establish the predictor’s value.

  2. The contribution of collaboration remains difficult to separate from attacker placement and generic role-based prompting. Tables 1 and 3 evaluate the combined framework, while section 3.2 specifies equal attacker counts without explaining Original’s placement rule. Fig 7’s “w/o Collab.” ablation partly addresses this, but its setting and whether identical attacker sets are retained are unclear; it also does not isolate the benefit of providing topology information to the planner. Matched comparisons with fixed attacker sets would clarify these contributions, while evaluation counts and uncertainty estimates would help assess the smaller gains.

Questions for authors
  1. How well does the predictor rank alternative attacker sets within a fixed graph and attacker budget, and how close is its selection to the best candidate?
  2. How does Original select its compromised agents and does Fig 7’s “w/o Collab.” comparison use identical attacker sets and tasks?
  3. How does the full predictor compare with a simple model trained only on the available structural features?
  4. What evaluation counts, independent repeat counts, and uncertainty estimates underlie Tables 1–3 and Fig 5–7?
  5. Do rankings predicted from basic-collaboration ASR remain consistent with the observed rankings after planner optimization?
  6. Does withholding edge information from the planner, while keeping the attacker set and execution graph fixed, reduce attack effectiveness?
7 comments
AuthorsOpenReview ↗

Response to Reviewer GLec (Part 1/3)

ear Reviewer GLec,

Thanks very much for your thoughtful feedback. We sincerely appreciate your insightful comments on the effectiveness of our vulnerability predictor, the contribution of structure-conditioned collaboration, and the rigor of our experimental evaluation. In response, we have conducted additional experiments and clarified our methodology to better validate the effectiveness and reliability of CIV-Amplify (Ours). Our detailed responses to your comments are provided below:

W1/Q1: The predictor's effectiveness in ranking and selecting attacker sets under fixed graphs and budgets remains unclear.

A: Thank you for raising this important concern. In the paper, Table 2 demonstrates the predictor's effectiveness in estimating attack risks, achieving an MAE of 0.0403, an R2R^2 of 0.8592, and a Spearman correlation (ρ\rho) of 0.9469. Figure 7 further provides downstream evidence: the full method achieves an ASR of 56.84%, outperforming density-based selection (50.52%) and random selection (46.32%).

To directly evaluate the predictor's ranking ability within a fixed graph and attacker budget, following your suggestion, we conducted additional experiments comparing it with a density-based baseline under Top-1 and Top-3 selection settings. As shown in Table 3, our predictor achieves ranking accuracies of 87.50% and 94.44% under Top-1 and Top-3 settings, respectively, outperforming the density-based baseline (75.00% and 86.11%) by 12.50 and 8.33. These results demonstrate that our predictor more accurately ranks alternative attacker sets and identifies highly vulnerable candidates under fixed graph and budget constraints.

Table 3: Top-1 and Top-3 Ranking Accuracy of the Predictor and Density-Based Baseline for Attacker-Set Selection.

MethodTop-1 (↑)Top-3 (↑)
Predictor87.5094.44
Density-based75.0086.11

Together, these results provide complementary evidence of the predictor's effectiveness in risk estimation, attacker-set ranking, and downstream selection.

W1/Q3: The full predictor’s added value over a simple model trained only on the available structural features remains unclear.

A: Thank you for this valuable suggestion. In the paper, Table 2 presents an ablation study demonstrating the contributions of Collaboration Risk, Individual Risk, and Count Risk to vulnerability prediction. Removing these components increases the MAE from 0.0403 to 0.0489, 0.0420, and 0.0422, respectively, confirming that they provide complementary information. However, these ablations do not directly establish the added value of the full predictor over a model using only structural features. To address this concern, following your suggestion, we conducted additional experiments comparing the full predictor with a structure-only baseline across three attack types.

Table 4: Vulnerability Prediction MAE (↓) of the Full Predictor and Structure-Only Baseline Across Attack Types.

Predictor VariantPrompt InjectionTool AttackMemory PoisoningOverall Avg.
Full predictor0.04030.05970.03840.0461
Structural features only0.05940.08700.08420.0769

As shown in Table 4, the full predictor consistently achieves lower MAE across all attack types, reducing the average MAE from 0.0769 to 0.0461 (40.05%). These results demonstrate that structural features alone are insufficient to fully capture attacker-subgraph vulnerability. By incorporating collaboration risk, individual risk, and count risk, our predictor captures complementary vulnerability information and provides more accurate attack-risk estimation.

AuthorsOpenReview ↗

Response to Reviewer GLec (Part 2/3)

W1/Q5: It remains unclear whether rankings learned from basic-collaboration ASR remain consistent after planner optimization.

Thank you for this valuable comment. As described in our methodology, the predictor aims to identify structurally vulnerable attacker subgraphs based on basic-collaboration ASR, thereby selecting promising attacker sets before planner optimization. As shown in Figure 2, this design is motivated by the observation that structural factors underlying Collaboration-Induced Vulnerability (CIV) play an important role in determining attacker-subgraph vulnerability. Although planner optimization further enhances attack effectiveness through structure-conditioned instructions, it does not change the underlying attacker subgraphs or their communication topologies. Therefore, the vulnerability rankings learned under basic collaboration can still provide useful guidance for subsequent attacker-set selection. To verify whether the rankings remain consistent after planner optimization, we conducted additional experiments comparing the predicted rankings under basic collaboration with the observed rankings after planner optimization, using Spearman's rank correlation ρ\rho and Top-3 overlap.

Table 5: Ranking Consistency Between Basic-Collaboration Predictions and Post-Optimization Attack Outcomes.

Attack TypeSpearman's ρ\rho (↑)Top-3 Overlap (↑)
Tool Attack0.900.87
Memory Poisoning0.970.93
Overall Avg.0.940.90

As shown in Table 5, the rankings remain highly consistent across both attack types. For Tool Attack, Spearman's ρ\rho and Top-3 overlap reach 0.90 and 0.87, respectively, while for Memory Poisoning, they reach 0.97 and 0.93. The overall averages are 0.94 and 0.90, respectively. These results demonstrate that vulnerability rankings learned from basic-collaboration ASR largely remain consistent after planner optimization. This validates our design choice of training the predictor using basic-collaboration data, avoiding the need for additional planner optimization and its associated API costs and computational overhead during dataset construction, while maintaining reliable attacker-subgraph selection.

W2/Q2: Original’s attacker placement rule and whether Fig. 7’s “w/o Collab.” comparison uses identical attacker sets and tasks are unclear, making collaboration gains difficult to isolate.

Thank you for raising this important clarification. As stated in Lines 392–393 of the paper, the Original baseline uses the same number of compromised agents as CIV-Amplify (Ours). Specifically, Original randomly selects the compromised agents, whereas CIV-Amplify (Ours) selects attacker agents based on the predicted vulnerability of candidate subgraphs.

Furthermore, to better demonstrate the effectiveness of CIV-Amplify (Ours), the "w/o Collab." comparison in Figure 7 applies the Original attack strategy to the attacker subgraphs selected by CIV-Amplify (Ours), without introducing additional collaborative instructions. Specifically, it retains exactly the same attacker sets, execution graphs, and evaluation tasks as the Full Method, while disabling the collaboration mechanism. This controlled comparison allows us to isolate the contribution of collaboration from attacker-subgraph selection and placement.

AuthorsOpenReview ↗

Response to Reviewer GLec (Part 3/3)

W2/Q6: The benefit of providing edge information to the planner remains unclear without a comparison that holds the attacker set and execution graph fixed.

A: Thank you for this insightful comment. In CIV-Amplify (Ours), the planner uses internal communication edges to generate topology-conditioned collaborative instructions. To isolate the contribution of edge information, following your suggestion, we conducted additional controlled experiments comparing the full method with two variants: one without role-specific instructions and the other without structure-conditioned planning (i.e., without providing edge information to the planner). Importantly, all three configurations use identical attacker sets, execution graphs, and evaluation tasks, ensuring a fair comparison.

Table 6: Ablation Study of ASR (↑) for Role-Specific Instructions and Structure-Conditioned Planning under Fixed Attacker Sets and Execution Graphs.

Attack TypeFull Methodw/o role-specific instructionw/o structure-conditioned planning
Prompt Injection25.7922.3721.58
Tool Attack46.3242.1139.21
Memory Poisoning36.5835.5329.74
Overall Avg.36.2333.3430.18

As shown in Table 6, removing structure-conditioned planning reduces the overall ASR from 36.23% to 30.18%, a decrease of 6.05, with consistent reductions across all three attack types. In particular, Tool Attack and Memory Poisoning decrease by 7.11 and 6.84, respectively. Removing role-specific instructions also reduces the overall ASR to 33.34%. These results demonstrate that both role-specific instructions and edge information contribute to attack effectiveness, with topology-conditioned planning providing substantial additional benefits beyond generic role-based collaboration.

W2/Q4: Evaluation counts, independent repeat counts, and uncertainty estimates for Tables 1–3 and Figs. 5–7 are needed to assess the reliability of the reported gains.

A: Thank you for this important comment. To clarify the evaluation scale, taking an 8-agent MAS as an example, each experimental setting involves 19 benchmark tasks with four interaction rounds, requiring 8×19×4=6088 \times 19 \times 4 = 608 agent accesses. Accordingly, a single run of Table 1 requires 58,368 agent accesses, Table 3 requires 72,960, Figure 5 requires 16,416, and Figure 6 requires 19,456 agent accesses. Due to the substantial API costs and execution time, we conducted two independent runs for the agent-based experiments.

For Table 2, predictor training uses the pre-generated dataset constructed as described in Section 2.2, without requiring additional agent interactions. We therefore evaluated the full predictor using five random seeds, obtaining MAE standard deviations of 0.0013, 0.0030, and 0.0020 for Prompt Injection, Tool Attack, and Memory Poisoning, respectively. Given the resource constraints and limited rebuttal timeline, we were unable to conduct additional repetitions. We sincerely hope for the reviewer's understanding and plan to perform more extensive evaluations with additional random seeds in future work.

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

This paper introduces the concept of Collaboration-Induced Vulnerability in multi-agent systems, aiming to reveal how the internal connectivity density among compromised agents and their outward propagation density toward benign agents jointly amplify existing multi-agent attacks. Building on this perspective, the authors design CIV-Amplify, a two-stage red-teaming framework that combines a graph neural network vulnerability predictor with a large language model collaborative instruction planner.

Critical strengths
  1. The paper establishes a complete pipeline from risk prediction and candidate search to collaborative attack generation by leveraging agent configurations, directed communication relationships, and attacker distributions.

  2. The main experiments span four victim models, four communication topologies, and three attack types, achieving consistent improvements in ASR.

Critical weaknesses
  1. Prior work such as MAStrike already identifies vulnerable agent coalitions and generates coordinated, role-aware attacks; against this backdrop, merely showing that connectivity and collaboration amplify attacks does not establish a distinct contribution. The experiments only compare against uncoordinated original attacks rather than closely related collaborative baselines like MAStrike or MASTER. Moreover, the ablation in Figure 7 does not isolate the benefit of structure-conditioned planning over an equally budgeted, topology-agnostic planner that generates role-specific instructions without structural inputs.

  2. The LLM planner only receives internal edges within the attacker subgraph as input while completely omitting outward edges to normal agents. Blind to the actual outward propagation topology, it produces generic role templates and even misidentifies a compromised attacker node as the victim target in the InjecAgent case study. Meanwhile, the ablation study shows that a training-free density heuristic combined with collaborative prompting already achieves nearly 70% of the total ASR gain without any GNN training. Consequently, of the paper's two core technical components, one is largely replaceable by a simple topological heuristic, while the other fails to utilize the complete local topology.

Questions for authors
  1. How does CIV-Amplify compare with MAStrike and MASTER under matched attack budgets?
  2. Why are outward edges omitted from the planner, and does it outperform an equally budgeted topology-agnostic planner?
  3. What additional benefit does the trained GNN provide over the training-free density heuristic?
7 comments
AuthorsOpenReview ↗

Response to Reviewer xnzg (Part 1/3)

Dear Reviewer xnzg,

Thanks very much for your thoughtful feedback and constructive suggestions. Your insightful comments have helped us further clarify the contributions of CIV-Amplify (Ours) and strengthen our experimental validation. Our point-by-point responses are provided below:

W1/Q1: The comparative effectiveness of CIV-Amplify against MAStrike and MASTER, and the contribution of topology-aware planning, require further validation.

Thank you for raising this important concern. As noted in the Introduction (Lines 052–053), MAStrike [1] studies explicit collusion by selecting multiple vulnerable agents and generating role-aware attacks. MASTER [2] also considers agent roles and communication topology when allocating attack tasks. Our contribution builds upon agent collaboration and topology-aware attack planning by further investigating the structural factors underlying Collaboration-Induced Vulnerability (CIV). Specifically, we examine how two distinct structural factors, namely connectivity among compromised agents and their outward connectivity to benign agents, affect attack amplification. We further leverage this structural vulnerability characterization to guide attacker-set selection and structure-conditioned collaborative planning.

Comparison with Existing Methods. To directly address the concern about comparisons with prior work, following your suggestion, we compare CIV-Amplify (Ours) with MAStrike and MASTER under matched attack budgets. As shown in Table 3, CIV-Amplify (Ours) achieves the highest ASR across all attack types, reaching 36.23% overall, outperforming MASTER and MAStrike by 6.58 and 7.72, respectively. Notably, it achieves gains of 8.43 and 8.16 over MASTER on tool attacks and memory poisoning, respectively. These results highlight the advantage of CIV-Amplify (Ours) in systematically exploiting collaboration-induced structural vulnerabilities. By jointly leveraging vulnerability-guided attacker-set selection and structure-conditioned collaborative planning, our method goes beyond conventional role-aware collaboration and topology-aware task allocation, achieving more effective attack amplification under the same attacker budget.

Table 3: ASR (↑) Comparison of CIV-Amplify (Ours) and Existing Collaborative Attack Methods Under Matched Attack Budgets.

Attack TypeOriginalMAStrikeMASTERCIV-Amplify (Ours)
Prompt Injection20.3921.8422.6325.79
Tool Attack32.2436.8437.8946.32
Memory Poisoning26.0526.8428.4236.58
Overall Avg.26.2328.5129.6536.23

Ablation of Collaborative Planning. To isolate the contribution of structural information, we compare our planner with an equally budgeted, topology-agnostic variant that retains role-specific instructions but receives no topology information. As shown in Table 4, removing role-specific instructions reduces the overall ASR from 36.23% to 33.34%, while removing structure-conditioned planning further reduces it to 30.18%. Both variants use identical attacker sets, execution graphs, and evaluation tasks. These results demonstrate that structure-conditioned planning provides substantial gains beyond role-based collaboration, highlighting the importance of exploiting structural vulnerabilities rather than merely coordinating attackers.

Table 4: Ablation Study of ASR (↑) for Role-Specific Instructions and Structure-Conditioned Planning under Fixed Attacker Sets and Execution Graphs.

Attack TypeFull Methodw/o role-specific instructionw/o structure-conditioned planning
Prompt Injection25.7922.3721.58
Tool Attack46.3242.1139.21
Memory Poisoning36.5835.5329.74
Overall Avg.36.2333.3430.18

Overall, these additional comparisons clarify that CIV-Amplify (Ours) is not simply another collaborative attack method. Its key contribution lies in characterizing structural vulnerabilities and exploiting them through vulnerability-guided attacker selection and topology-conditioned collaborative planning.

References

[1] Xu C, Chen Z, Zhang J, et al. MAStrike: Shapley-Guided Collusive Red-Teaming on Multi-Agent Systems[J]. arXiv preprint arXiv:2606.12918, 2026.

[2] Zhu Y, Zhang C, Shi X, et al. MASTER: Multi-Agent Security Through Exploration of Roles and Topological Structures-A Comprehensive Framework[C]//EMNLP (Findings). 2025: 16895-16921.

AuthorsOpenReview ↗

Response to Reviewer xnzg (Part 2/3)

W2/Q2: The planner omits outward edges to normal agents, and its advantage over an equally budgeted, topology-agnostic planner remains unclear.

A: Thank you for this thoughtful question. We clarify that internal and outward edges serve different purposes in CIV-Amplify (Ours). Internal edges capture communication among compromised agents and are provided to the planner to guide role assignment and collaborative strategy generation. In contrast, outward edges describe connections between compromised and benign agents. Rather than explicitly providing these edges to the planner for propagation-path planning, we allow every compromised agent to leverage its own outward connections to propagate attack-related information to its benign neighbors through the communication protocol. Meanwhile, the planner dynamically assigns complementary roles, including propagation-related roles, based on the selected subgraph's internal structure and attack objective. This design enables all compromised agents to participate in outward attack-related information propagation through their respective connections, while the planner coordinates their internal collaboration and assigns complementary propagation responsibilities.

To further investigate whether explicitly providing outward edges to the planner improves collaborative planning, we conduct an additional ablation by including them as planner inputs. As shown in Table 5, this modification reduces the overall ASR from 36.23% to 34.47%, with decreases across all three attack types. This indicates that explicitly incorporating outward edges into planning does not provide additional benefits in our setting. Instead, our design allows the planner to focus on coordinating compromised agents while leveraging their outward connections for attack-related information propagation. This separation enables the framework to fully exploit internal connectivity for collaborative coordination while allowing all compromised agents to independently propagate attack-related information through their outward connections. It also avoids introducing potentially redundant structural information into the planner, reducing unnecessary planning complexity.

Table 5: Effect of Incorporating Outward Edges into Collaborative Planning on ASR (↑).

MethodPrompt InjectionTool AttackMemory PoisoningOverall Avg.
CIV-Amplify (Ours)25.7946.3236.5836.23
+Outward edges25.2644.7433.4234.47

We further compare our planner with an equally budgeted, topology-agnostic variant. As shown in Table 4, removing structural information reduces the overall ASR from 36.23% to 30.18%. This demonstrates that internal connectivity provides important guidance for collaborative planning. Together, these results support our design choice of using internal edges to guide collaboration while handling outward propagation through the communication protocol.

Regarding the concerns about generic role templates and the InjecAgent case study, we clarify that the role descriptions (e.g., coordination, independent rationale, propagation, and check) in the prompt are illustrative examples rather than predefined templates. The planner dynamically determines and assigns complementary roles based on the internal connectivity of the selected attacker subgraph and the attack objective, considering each agent's structural position and potential contribution to collaboration. However, we acknowledge that Figure 11 does not sufficiently illustrate how the injected instruction is propagated from compromised agents to benign agents, which are the actual attack targets. This may misleadingly suggest that a compromised agent is treated as the victim. We appreciate the reviewer for pointing this out and will replace Figure 11 with a more representative example that better demonstrates the dynamic role assignment and subsequent propagation to benign agents.

AuthorsOpenReview ↗

Response to Reviewer xnzg (Part 3/3)

W2/Q3: The added value of the GNN predictor over the density heuristic remains unclear.

Thank you for your careful attention to our GNN predictor. Our motivation for introducing the GNN predictor originates from our analysis of Collaboration-Induced Vulnerability (CIV), which reveals that both the connectivity among compromised agents and their outward connectivity to benign agents influence attack amplification. The density heuristic exploits these structural observations by selecting attacker subgraphs that follow the identified density trends, making it an effective training-free baseline.

However, density captures only aggregate connectivity and cannot fully characterize the underlying directed structures. Different attacker subgraphs may exhibit similar internal and outward densities but differ in their edge distributions, node connectivity patterns, and potential for collaboration. Such structural differences may lead to different levels of attack vulnerability, which cannot be adequately distinguished by density alone. We therefore introduce a directed GNN predictor to capture these finer-grained structural patterns and provide more expressive vulnerability estimates for attacker-set selection.

The experimental results support this motivation. As shown in Figure 7, replacing the density heuristic with our trained predictor increases ASR from 50.52% to 56.84%, yielding an additional gain of 6.32 (12.51% relative improvement) under the same collaborative planning setting. These results demonstrate that, while the density heuristic captures the overall structural trends underlying CIV, the GNN predictor provides additional modeling capacity to distinguish structurally different attacker subgraphs and identify more vulnerable candidates.

Reviewer_kagvOpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies how connections between compromised agents amplify attacks in LLM multi-agent systems. It defines Collaboration-Induced Vulnerability through coordination among attackers and propagation to normal agents. CIV-Amplify combines a directed-GNN predictor with an LLM planner to choose attacker subgraphs and coordinate their behavior. Across four topologies, four backbones, and three attack families, average attack success rises from 26.58% to 36.27%. The experiments also cover existing safeguards, larger graphs, and component ablations.

Critical strengths
  • Separating attacker coordination from outward propagation gives a useful way to study structural risk. Figure 9 supports the relevance of both factors across three attack families.
  • The predictor makes the structural analysis practical. Figure 8b shows that guided subgraph selection is much faster than exhaustive search.
  • Table 1 covers several topologies, backbones, and attacks rather than one favorable setup. The safeguard tests and component ablations provide useful additional checks.
Critical weaknesses
  • Figure 7 separates selection and planning, but does not compare basic collaboration with structure-conditioned collaboration at inference. This makes the planner-specific contribution to Table 1's gains hard to judge.
  • The predictor learns from eight-agent configurations run with GPT-4o-mini, then transfers to other backbones and larger graphs (Appendix B). Its ranking accuracy on those new settings and run-to-run variation are not reported.
  • Gains vary across settings, including no overall improvement for DeepSeek-V3 on chains (Table 1). Figure 6 pools safeguard results without a per-attack or per-backbone breakdown, so the generality of those results is unclear.
Questions for authors
  • What collaboration does the Original baseline allow? Could you compare basic and structure-conditioned collaboration while keeping subgraph selection fixed?
  • How well does the predictor rank candidate subgraphs on unseen backbones and 80-agent graphs? Comparing its choices with exhaustive selection on small graphs would help.
  • In Figure 6, what can each safeguard inspect? Please clarify whether its access includes only inter-agent messages or also system prompts and planner-generated instructions.
2 comments
AuthorsOpenReview ↗

Response to Reviewer kagv (Part 1/2)

Dear Reviewer kagv:

Thanks for your careful review and constructive feedback. We greatly value your positive comments on our work and the thoughtful suggestions you provided, which have helped us improve both the technical clarity and overall presentation of the paper. We respond to each comment in detail below:

W1/Q1:The Original baseline’s collaboration mechanism is unclear, and a comparison of basic versus structure-conditioned collaboration with subgraph selection held fixed is needed to isolate the planner’s contribution.

A: Thank you for this valuable feedback. As described in Line 452 of the paper, the w/o Collab variant removes the collaborative optimization module while retaining the original interaction instructions used in the Original baseline. The Original baseline allows basic message-based collaboration: attackers are encouraged to consider other agents' perspectives and can exchange messages through the communication protocol, enabling implicit coordination. However, it does not explicitly define joint strategies, differentiated attacker roles, or topology-conditioned planning.

To further distinguish the contributions of basic collaboration and structure-conditioned collaborative optimization, following your suggestion, we introduce a three-way comparison in Table 3. Specifically, we first use the CIV-Amplify predictor to identify vulnerable attacker subgraphs and then keep the selected attacker sets, execution graphs, and evaluation tasks fixed across all three variants, changing only the collaboration mechanism. The results show that basic collaboration achieves an overall ASR of 33.34%, compared with 30.09% without collaboration. Incorporating structure-conditioned collaborative optimization (Ours) further increases the ASR to 36.23%. This consistent improvement across all three attack types demonstrates that, even with the attacker subgraphs fixed, topology-conditioned collaborative optimization provides additional gains beyond basic message-based collaboration.

Table 3: Ablation Study of ASR (↑) for Basic Collaboration and Structure-Conditioned Collaborative Optimization with Fixed Attacker Sets and Execution Graphs.

Attack TypeCIV-Amplify (Ours)Basic CollaborationNo Collaboration
Prompt Injection25.7922.3721.05
Tool Attack46.3242.1135.53
Memory Poisoning36.5835.5333.68
Overall Avg.36.2333.3430.09

W2/Q2: The predictor’s ranking accuracy and run-to-run variability on unseen backbones and 80-agent graphs remain unclear, and comparison with exhaustive subgraph selection on small graphs is needed.

A: Thank you for this thoughtful question. To address the concern regarding ranking accuracy under new settings, following your suggestion, we conducted additional experiments using different backbone models (GPT-4o-mini and DeepSeek-V3) and agent configurations (8, 20, 40, and 80 agents). As shown in Table 4, the predictor maintains consistently high ranking accuracy across different backbones and agent scales, suggesting good transferability and scalability.

Table 4: Predictor Ranking Accuracy (↑) Across Different Backbone Models and Multi-Agent System Scales.

Backbones8 Agent (Exhaustive Search)20 Agent (Guided Search)40 Agent (Guided Search)80 Agent (Guided Search)
GPT-4o-mini94.4490.2887.5084.72
DeepSeek-V393.0688.8986.1184.72

Regarding run-to-run variation, for the 8-agent exhaustive search setting, there are 24 candidate subgraphs. To ensure a fair comparison, we also restrict the guided search methods to the same 24 candidate subgraphs. Evaluating all configurations in Table 4 alone requires a total of 539,904 agent accesses. Due to the substantial API costs and execution time, as well as the limited rebuttal timeline, we were unable to conduct multiple independent runs for all configurations. We sincerely hope for your understanding regarding this limitation. We plan to conduct more extensive repeated experiments with different random seeds in future work.

AuthorsOpenReview ↗

Response to Reviewer kagv (Part 2/2)

W3/Q3: The generality of safeguard effectiveness across attack types and backbones, and the scope of safeguard information access, remain unclear.

A: Thank you for this insightful comment. In Figure 6, we evaluated the effectiveness of different MAS safeguards (BlindGuard, Dominant, and G-SafeGuard) under the Prompt Injection attack scenario. For a fair comparison, all safeguards were restricted to inspecting inter-agent messages only.

Following your suggestion, we further conducted experiments to provide a more fine-grained evaluation of different safeguards across attack types and backbone models. Specifically, we evaluated three attack types (Prompt Injection, Tool Attack, and Memory Poisoning) using two backbones (GPT-4o-mini and DeepSeek-V3), and reported the results separately for each attack type and safeguard in Table 5.

Table 5: Per-Attack ASR (↑) of Different Safeguards Across Backbone Models and Attack Settings.

BackboneDefenseSettingPrompt InjectionTool AttackMemory PoisoningOverall Avg.
GPT-4o-miniNo DefenseOriginal20.3932.2426.0526.23
GPT-4o-miniNo Defense+Ours25.7946.3236.5836.23
GPT-4o-miniBlindGuardOriginal15.784.2123.6814.56
GPT-4o-miniBlindGuard+Ours20.0018.4227.6322.02
GPT-4o-miniDominantOriginal17.896.5825.2616.58
GPT-4o-miniDominant+Ours18.9419.7428.9522.54
GPT-4o-miniG-SafeGuardOriginal10.533.4222.1112.02
GPT-4o-miniG-SafeGuard+Ours15.7816.3227.3719.82
DeepSeek-V3No DefenseOriginal15.1326.7143.0328.29
DeepSeek-V3No Defense+Ours21.8427.1159.4736.14
DeepSeek-V3BlindGuardOriginal12.639.2137.1119.65
DeepSeek-V3BlindGuard+Ours13.6813.1639.2122.02
DeepSeek-V3DominantOriginal11.5810.0038.1619.91
DeepSeek-V3Dominant+Ours11.5814.4739.7421.93
DeepSeek-V3G-SafeGuardOriginal5.267.8933.4215.52
DeepSeek-V3G-SafeGuard+Ours8.4214.4736.8419.91

The additional results provide a more fine-grained understanding of safeguard performance across different settings and demonstrate that the effects of CIV-Amplify (Ours) vary across attack types and backbone models. Nevertheless, the overall results provide further evidence of its effectiveness across the evaluated settings.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Responses To All Reviewers

Dear Reviewers and Area Chairs,

We sincerely thank all reviewers for their constructive feedback and for recognizing our characterization of Collaboration-Induced Vulnerability (CIV) and the breadth of our evaluation (Reviewers kagv, xnzg, and GLec). We particularly appreciate the recognition of our distinction between compromised-agent coordination and outward propagation to benign agents (Reviewers kagv and GLec).

Our additional experiments and clarifications address baseline fairness, the value and cost of individual components, and the reliability and scope of our findings. The key additions are summarized below.

  • We compare CIV-Amplify with MAStrike and MASTER under matched online budgets and report defense results separately across attack types and backbones.

  • With attacker sets and execution graphs fixed, we ablate basic collaboration, role assignment, structure-conditioned planning, and outward-edge information to isolate their respective contributions.

  • We compare the GNN predictor with density-based and structure-only alternatives, report data and computational costs, assess sensitivity across seeds and training data, and test transfer to an unseen backbone and larger graphs.

  • We explain how our controlled evaluation isolates the effect of collaboration structure. We will also explicitly delimit our claims and discuss validation in production-style workflows as future work.

We sincerely thank all reviewers again for their valuable feedback. We believe that these additional experiments and clarifications strengthen the empirical rigor of our work, better delineate its scope and core contributions, and improve the overall completeness and clarity of the manuscript.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.