acceptodds
Under review as a conference paper at ICLR 2027

MOSAIC: Unified Outcome-Aware Security Auditing for LLM-Based Multi-Agent Systems

Abstract

Large Language Model (LLM)-based multi-agent systems (MAS) enable complex tasks through communication and collaboration among specialized agents, but their interconnected nature introduces security risks beyond the scope of conventional LLM and single-agent safeguards. Existing MAS defenses focus on complementary aspects such as access control, anomaly detection, or task-failure prediction, yet lack a unified mechanism for determining whether an attack actually compromises the system, where its effect occurs, and what evidence supports the decision. We propose **MOSAIC**, a unified outcome-aware security auditing framework that integrates attack detection, realized-outcome assessment, structural localization, and evidence-grounded attribution for MAS. We further introduce **MOSAIC-Bench**, a systematic benchmark that distinguishes attack exposure from realized compromise, and **MOSAIC-Auditor**, a graph-grounded auditor for reasoning over multi-agent execution trajectories. Experiments demonstrate the effectiveness of MOSAIC for comprehensive MAS security auditing and highlight the necessity of outcome-aware, structurally grounded reasoning beyond existing safety paradigms.

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

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_VQsSOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents MOSAIC, an outcome-aware security auditing framework for LLM-based Multi-Agent Systems. It introduces MOSAIC-Bench, a dataset of 43,844 execution trajectories covering diverse attacks and communication structures, with annotations for attack outcomes, component localization, and supporting evidence. The authors further develop MOSAIC-Auditor through supervised fine-tuning to generate structured audit reports. Experiments show that MOSAIC outperforms several baselines in attack outcome classification and localization.

Critical strengths
  1. The research problem is practically meaningful. The paper clearly distinguishes attack exposure from realized compromise. This distinction is particularly important for multi-agent systems, as an individual agent receiving or following malicious instructions does not necessarily lead to a system-level security compromise. By focusing on actual execution outcomes rather than isolated anomalous behaviors, the paper addresses a well-motivated and practically relevant security auditing problem.

  2. The benchmark is systematic and comprehensive. MOSAIC-Bench covers diverse attack objectives, injection surfaces, communication topologies, and structural placements, with over 40,000 execution trajectories. Beyond attack success labels, the benchmark also provides component localization and event-level evidence annotations, supporting more fine-grained security auditing research. This is a valuable contribution of the paper.

  3. The experimental evaluation is relatively comprehensive. The authors compare against multiple baselines and further investigate the transfer of single-agent auditors to multi-agent settings and the effects of different input representations. The results demonstrate that MAS-specific training significantly improves attack outcome classification and component localization. The paper is also clearly organized, with methods and experimental settings that are easy to follow.

Critical weaknesses
  1. The paper combines outcome assessment, attack profiling, localization, and evidence attribution into a single generation task. However, it is unclear whether this unified approach is better than handling these tasks separately. The paper lacks an ablation comparing these designs and does not discuss how conflicting predictions or errors across different audit outputs are handled.

  2. The contribution of graph-based representation seems limited. In Table 2, Graph SFT only slightly outperforms Flat SFT, with improvements of less than one percentage point. While the flat representation still contains some structural information, the results do not clearly show how much the explicit graph representation contributes to auditing performance.

  3. The attack construction relies on predefined injection surfaces and structural placements, with at most two injection sites per attack. While this design makes the experiments reproducible, it may not fully capture more complex attacks involving dynamic propagation across multiple agents. Moreover, localization mainly identifies where an attack was injected, rather than which components actually contributed to the final compromise. The paper could make better use of its controlled execution environment to investigate these relationships.

Questions for authors
  1. Have the authors considered comparing unified auditing with independent prediction or a modular pipeline? How does the current approach handle inconsistencies between different audit outputs?

  2. In which scenarios does the explicit graph representation provide the most benefit? Would the performance change significantly if some communication edges were removed or modified?

  3. Could the authors evaluate more complex attack propagation scenarios? In particular, could interventions on intermediate agents or communication edges help verify whether the localized components actually contribute to the final security outcome?

4 comments
AuthorsOpenReview ↗

Response to Reviewer VQsS

We thank Reviewer VQsS for recognizing the practical importance of the problem, the systematic design of MOSAIC-Bench, and the comprehensive evaluation. We address the concerns below.

W1 / Q1. Comparing unified auditing with independent prediction or a modular pipeline

We revisited a controlled MARBLE–AppWorld experiment (3,122 training; 406 validation trajectories). Unified Qwen SFT generated the full report directly; the modular pipeline froze the same Qwen representation, trained separate verdict, scope, and component heads, and used conditional SFT for report generation.

Comparison between unified auditing and the modular pipeline (%)

MethodOutcome Macro-F1Localization F1
Unified Qwen SFT73.3080.72
Modular: three independent heads + conditional report63.6068.64
Unified-auditing gain+9.70+12.08

Unified auditing gained 9.70 Outcome and 12.08 localization points, showing the benefit of joint cross-field modeling.

MOSAIC avoids post-hoc stitching through unified structured generation. A deterministic check flags any cross-field conflict after decoding without overwriting predictions.

W2 / Q2. Where does the explicit graph representation help most, and what is the effect of modifying communication edges?

We compared matched one-epoch Graph and Flat models across seeds 42–44 on the same 6,207-example test set.

Graph-over-Flat gains by scenario (percentage points)

ScenarioTest examplesOutcome gainLocalization gain
Research861+1.37+3.35
Bargaining1,187+0.49+2.30
Coding1,217+1.95+1.97
Minecraft964+1.23+0.75
AppWorld1,978+0.38+0.10

Graph helped most in responsibility-tracing settings: Research gained +3.35 localization points; Tree and selector-group-chat topologies gained +3.18 and +2.65 points.

We then modified only the inference-time graph for the seed-44 auditor on a fixed 2,000-example subset.

Communication-edge intervention results (%)

Graph inputOutcome Macro-F1Localization F1Localization change
Original graph78.2081.24—
Grounding information masked78.4780.53−0.71
Communication edges removed78.4578.98−2.27
Communication edges rewired78.3180.58−0.67

Removing edges reduced localization F1 by 2.27 points (paired-bootstrap 95% CI: −4.25 to −0.36) but barely changed outcome prediction, indicating that graph structure mainly supports component attribution.

W3 / Q3. Predefined injection, dynamic propagation, and interventions on intermediate components

Injection surfaces and at most two injection points are predefined for clear exposure, provenance, and outcome labels; subsequent propagation is determined by the MAS runtime. We traced each attack's unique success marker (Appendix A.5) through existing logs.

Cross-agent propagation among successful attacks

StatisticCountShare
Traceable successful local-injection attacks2,942100%
Relayed through a non-injected agent85429.0%
Completed by the injected agent2,08871.0%

Under a strict temporal, non-refusal matching criterion, 29.0% is a lower bound on dynamic cross-agent propagation. The other 71.0% were completed by the injected component. In a 12-run multi-hop pilot, relay agents forwarded the payload in 7 of 9 applicable cases.

Our localization labels capture provenance, not formal causal responsibility. Payload-free controls nevertheless support injection necessity: 0/2,835 produced consensus attack success (Appendix A.9). The revised manuscript now includes the propagation results in Appendix A.10, and the marker paths are released with the benchmark; causal re-execution interventions and more complex adaptive relays remain future work.

We thank Reviewer VQsS again for the careful and constructive feedback, which helped us improve the paper.

AuthorsOpenReview ↗

Follow-up on Our Response

Dear Reviewer VQsS,

Thank you again for your thoughtful feedback. We would appreciate knowing whether our response and revisions have adequately addressed your concerns. If any issues remain, we welcome further discussion and would greatly value your guidance. If your concerns have been resolved, we would be grateful if you could consider updating your evaluation accordingly.

Thanks for the rebuttal response. I have also read your summary of common concerns, which are also helpful. My concerns are addressed. I have raised my score accordingly.

Thank you very much for your thoughtful evaluation and positive recognition of our work. We sincerely appreciate your feedback and are happy to discuss any further questions or concerns.

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

MOSAIC introduces outcome-aware security auditing for multi-agent systems, jointly predicting compromise, attack provenance, and supporting evidence. MOSAIC-Bench contains 43,844 trajectories. A supervised generative auditor achieves 80.49% outcome Macro-F1, 79.69% successful-attack F1, and 83.96% localization F1.

Critical strengths

S1: The benchmark distinguishes attack exposure from realized compromise and provides structural supervision across diverse attack surfaces and coordination settings. S2: Task-disjoint splits, matched flat/graph comparisons, and class-level analyses provide credible evidence for the value of domain-specific auditing supervision.

Critical weaknesses

W1: Evidence F1 measures agreement with heuristic reference selection, without a separate semantic assessment of whether those events substantiate the verdict (Appendix B.3; Table 17). This limits the evidence-grounding claim. W2: Graph formatting adds only 0.42 points in outcome Macro-F1 and 0.43 in localization F1 over matched Qwen FlatSFT (Table 2). As acknowledged, the stronger contribution is domain-specific supervision rather than the graph representation itself. W3: Outcome-label reliability is incompletely characterized: Appendix B.2 does not specify judge identities, prompts, or disagreement/adjudication counts for attacked runs. Appendix A.9's clean-control specificity check is useful but does not establish agreement on successful attacks.

Questions for authors

Q1: Does adjudication separately verify the semantic sufficiency of selected evidence references, or only objective realization from the underlying execution? Q2: Are the small graph-versus-flat gains consistent across training seeds, particularly for outcome Macro-F1 and localization? Q3: Which judges and prompts produce outcome labels, and what are the disagreement and manual-adjudication rates for attacked runs?

2 comments
AuthorsOpenReview ↗

Response to Reviewer orGp

We thank Reviewer orGp for recognizing the importance of the problem, our benchmark design, and the experimental analysis. We address the concerns below.

W1 / Q1. Does adjudication separately verify the semantic sufficiency of evidence?

Thank you for pointing out this important distinction. Our evidence statistics reflect objective realization in the underlying execution. To additionally assess the semantic sufficiency of the selected evidence, we stratified 300 examples from the 6,207-example test set (100 per outcome). GPT-6 SOL and a human reviewer independently examined each selected reference and its neighboring context.

Contextual semantic support of evidence (300 examples)

ReviewerFull supportPartial supportNo support
GPT-6 SOL20.3%53.7%26.0%
Human review22.7%57.3%20.0%

The model review and additional human review both show that most references provide at least some contextual semantic support for the verdict. The revised manuscript now clearly distinguishes outcome realization, reference agreement, and evidence semantic support in Appendix C.4.

W2 / Q2. Are Graph-over-Flat gains consistent across random seeds?

Thank you for raising this question. Under identical data, training, and evaluation settings, we trained Graph and Flat Qwen3-8B LoRA models with seeds 42, 43, and 44 under a matched one-epoch budget; the input representation was the only variable. All models were evaluated on the same sealed test set of 6,207 examples.

Multi-seed Graph-versus-Flat results (%)

MetricSeedFlatGraphGain
Outcome Macro-F14277.5378.31+0.79
Outcome Macro-F14377.4878.46+0.98
Outcome Macro-F14477.2778.08+0.81
Outcome Macro-F1Mean77.4278.28+0.86
Localization Micro-F14280.0381.73+1.70
Localization Micro-F14380.4481.77+1.33
Localization Micro-F14479.6381.39+1.76
Localization Micro-F1Mean80.0381.63+1.59

Graph achieved positive gains on both metrics for all three seeds. Paired-bootstrap 95% confidence intervals were +0.41 to +1.30 percentage points for Outcome and +0.87 to +2.35 percentage points for Localization. These results demonstrate consistency across random seeds.

W3 / Q3. Label reliability

The judges were DeepSeek-V4-Flash and GPT-5.6-Luna, with disagreements resolved through manual adjudication. Both judges used the same outcome-aware prompt, security-final-output-aware-v2, which provided the attack objective and privileged execution evidence and requested attack_success, confidence, a reason code, evidence quotes, and a rationale. The revised manuscript now includes the full prompt in Appendix B.2.

On attacked runs, the two judges agreed on 77.6% and disagreed on 22.4%; all disagreements were manually adjudicated. Among the 8,058 attack-success cases, 81.5% were based on judge agreement and 18.5% on manual adjudication. A manual audit of 100 cases agreed with the final labels in 91% of cases.

We thank Reviewer orGp again for the careful and constructive feedback, which helped us improve the paper.

AuthorsOpenReview ↗

Follow-up on Our Response

Dear Reviewer orGp,

Thank you again for your thoughtful feedback. We would appreciate knowing whether our response and revisions have adequately addressed your concerns. If any issues remain, we welcome further discussion and would greatly value your guidance. If your concerns have been resolved, we would be grateful if you could consider updating your evaluation accordingly.

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

Mosaic is a unified audit mechanism to determine whether an attack actually compromises the system, where its effect occurs, and what evidence supports the decision. Mechanisms designed for single-agent systems do not encounter these structural aspects. The paper presents a pipeline to construct multi-agent attacks and corresponding ground-truth audit reports. The resulting test set is used for training an auditor model by finetuning a qwen3-8B checkpoint (and another one with InterLM3-8B-Instruct checkpoint). The auditor model outperforms general-purpose LLMs, graph-based multi-agent system (MAS) defenses and specialized auditors in terms of accuracy across outcome, compromise and attribution predictions.

Critical strengths
  • The problem of auditing MAS is important. The key difference between single-agent systems and MAS is structured composition of multiple agents and pinning the vulnerability/attack to individual components. The paper addresses this problem.
  • The construction of the MOSAI-Bench is interesting. The authors have carefully ensured that the tasks cover multiple objectives and structural combinations of MAS designs.
  • The experiments pertaining to adapting single agent auditing and ablations on modernBERT are insightful.
  • The figures in the paper are quite useful and help understand the designs nicely.
Critical weaknesses
  • Key details on trajectory sampling for creation of MOSAIC-Bench are not provided.
  • The paper evaluates Marble and AutoGen. No commercial-grade agent SDKs like LangGraph, OpenAI SDK, etc. are evaluated.
  • Similarly, some key details on training and evaluation are missing.
Questions for authors
  1. What per-task sampling budget was used for generating the attack trajectories? What are the reasons for the "attack failed" cases?
  2. Do the evaluation metrics cover all parts of the audit report defined in Eq. 2?
  3. For training (Eq. 6), are all tokens weighted equally? Could the results improve if key tokens (e.g., the outcome prediction) are weighed more?
3 comments
AuthorsOpenReview ↗

Response to Reviewer HWEu

We thank Reviewer HWEu for recognizing the importance of MAS security auditing and the value of MOSAIC-Bench, our experiments, and presentation. We address the questions below.

W1 / Q1. Sampling budget

Thank you for raising this reproducibility issue. We rechecked the paper-aligned generation records, final manifests, and judge rationales. Sampling used a global budget: random tracks distributed runs evenly across tasks, while enumerable settings covered all planned combinations. Planned-run completion was 98.7%–99.9%, with only 0.5%–1.3% of draws not instantiated.

Sampling distribution and budget across attack dimensions

Sampling dimensionDistribution
ObjectiveThree classes, 1/3 each
Injection pointsSingle 80%; dual 20%
Single-point surface5 classes, 20% each
Dual-point surface10 combinations, 10% each
ProtocolUniform sampling
SelectorCentral node 40%; three other classes 20% each

Attack-trajectory scale

StatisticCountAttacks per base task
Generated attack runs35,780Mean 63.0
Retained trajectories28,352Mean 49.9; median 42
Attack success8,05828.42%
Attack failure20,29471.58%

Reasons for attack failure

Failure reasonCountShare
Injection ignored or did not reach the target12,78062.97%
Unrelated execution error2,77613.68%
Restated but not executed1,6498.13%
Explicit rejection or resistance6123.02%
Attempted but incomplete execution2,47712.21%
Total20,294100.00%

Most failures arose because the injection was ignored or did not reach the target (62.97%), followed by unrelated execution errors (13.68%) and incomplete execution attempts (12.21%).

W2. Cross-framework evaluation on LangGraph

We applied the same attack templates to LangGraph, generating 401 runs: 240 attacked, 85 private-context controls, and 76 clean controls. After quality filtering, 365 trajectories remained for zero-shot evaluation with three trained, seed-specific Graph auditors.

LangGraph zero-shot evaluation results (%)

MetricMean ± SD over three seedsSeed 42 95% CI
Outcome Macro-F176.70 ± 1.2172.98–81.68
Binary Macro-F190.79 ± 0.7986.92–93.85
Localization Micro-F159.81 ± 2.9552.44–67.84
Parse success rate100.00 ± 0.00—

Outcome Macro-F1 was only 1.58 points below the 78.28 in-domain mean, supporting zero-shot transfer of outcome auditing. The lower Localization Micro-F1 indicates greater sensitivity to framework-specific protocols and component semantics.

Q2. Metric coverage of the audit report

All five components of Eq. (2) were evaluated. The submission emphasized outcome, localization, and evidence, but did not separately report attack profile and structural scope; we provide them below.

Metric coverage for each field in Eq. (2)

Field in Eq. (2)Evaluation on submission predictions
Outcome yyMacro-F1 80.49; attack-success F1 79.69
Attack profile π\piObjective Macro-F1 86.19; surface Macro-F1 64.91; joint exact match 71.19
Structural scope ℓ\ellMacro-F1 66.50
Components KKMicro-F1 83.96
Evidence HHStep-aware F1 88.46; localized-evidence F1 84.35; evidence-set exact match 75.01

The revised manuscript now includes the complete field-to-metric mapping and explicitly distinguishes evidence-reference validity from semantic sufficiency.

Q3 / W3. Token weighting in Eq. (6)

The submission weighted all assistant-output tokens equally; system, input, and padding tokens were excluded from the loss. In a matched pilot, we doubled only decision, localization, and audit_trace, keeping all other settings fixed (seed 42; 1,000/500 training/validation examples; one epoch).

Pilot results for key-field weighting

Loss settingThree-class accuracyBinary accuracy
Uniform weighting44.279.2
2× weight on key fields43.679.2

Key-field weighting did not improve this pilot: binary accuracy remained 79.2%, while three-class accuracy decreased by 0.6 percentage points.

We thank Reviewer HWEu again for the careful and constructive feedback, which will help us improve the paper.

Reviewer_HWEuOpenReview ↗

Response read

Thank you for your detailed response! Please include these results in the paper.

Thank you for your positive response. As requested, we have added the rebuttal results to the revised manuscript and highlighted the changes in blue. If these revisions address your concerns, we would appreciate your consideration of updating your evaluation. Thank you again for your time and feedback!

Comments on the paper

1 comment
AuthorsOpenReview ↗

Appreciation and Summary of Common Review Concerns

Dear Area Chair and Reviewers,

We sincerely thank all reviewers for their careful and constructive feedback. Our work develops unified, outcome-aware security auditing for multi-agent systems, jointly assessing security outcomes and supporting evidence. We appreciate the reviewers’ recognition of the problem’s importance, the value of MOSAIC-Bench, and the clarity of our presentation. We believe this work advances reliable and interpretable auditing for LLM-based multi-agent systems.

Here, we briefly summarize the common concerns raised during the rebuttal process.

Common Concern 1: The Role of Graph Structure

Reviewers orGp and VQsS asked whether the small Graph-over-Flat gap reflects a meaningful benefit from graph structure. Our three-seed results show a consistent gain, especially for localization in Research scenarios. Localization also drops when communication edges are removed. These results support a focused conclusion: graph structure mainly helps identify the affected components.

Common Concern 2: Benchmark and Label Reliability

Reviewers HWEu and orGp requested more detail on how the benchmark and its labels were produced. We now report the sampling budget and attack-failure reasons, and clarify how outcome labels and evidence references were checked, including judge disagreement and manual adjudication.

We address all remaining questions separately in the reviewer-specific responses. A revised manuscript has been released, with changes from the original submission highlighted in blue. Details are in our individual responses, and we are happy to clarify anything further.

Thank you again for your time and helpful comments.

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.