acceptodds
Under review as a conference paper at ICLR 2027

Shattered Intent: Payload Sharding Attacks on Multi-Agent LLM Systems

Abstract

Multi-agent systems (MAS) decompose complex tasks across specialized agents, with each agent accessing only the inputs assigned to its role. We show that this structure creates a vulnerability we term payload sharding, where a malicious payload is split into individually benign shards distributed across agents. We instantiate the attack as Mosaic, which decomposes a payload along intent, logic, and target boundaries aligned with the plan-code-execute workflow of code-generating MAS. Shards are synthesized by an iterative procedure that generates a cover context, decouples logic from sensitive parameters, and verifies via LLM auditing that each shard appears benign in isolation. Mosaic achieves ASR across five LLMs and five MAS architectures, exceeding the strongest prior attack by . Under defenses, it retains ASR against LlamaFirewall and against LLM-based cross-artifact auditing.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM safety, Multi-agent system

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 2 reviews, above 9% of the others.
Reviewer_nZR3OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces MOSAIC, an attack framework that splits a malicious operation across intent, code logic, and target configuration artifacts for runtime reassembly via normal multi-agent collaboration, exposing the limitations of local safety inspections against cross-artifact malicious compositions.

Critical strengths
  1. The paper astutely identifies that the very mechanisms modern MAS frameworks rely on to enhance engineering capabilities, paradoxically create a structural blind spot for safety defenses.

  2. The paper organizes code parameterization, configuration decoupling, and software-engineering task camouflage into an executable attack pipeline, leveraging ASCR for automated shard generation, auditing, and iterative refinement.

Critical weaknesses
  1. The paper attributes its high ASR to the premise that the complete malicious operation never surfaces in any single agent's context. However, the comparisons in Tables 1–2 simultaneously change the engineering narrative, code organization, parameter placement, and payload presentation, making it impossible to distinguish whether the gains arise from cross-agent information isolation or simply from code parameterization and semantic camouflage. Table 4 only reports offline detection rates from a GPT-4o judge rather than the end-to-end ASR of these variants, and it lacks a controlled comparison of the exact same generated artifacts across single-agent systems, fully shared-view MAS, and local-view MAS.

  2. The paper misstates several key differences from prior work, making it difficult to judge its incremental contribution. Appendix B.1 claims that MAS-CFH requires runtime intervention, though its threat model also covers passive attacks via untrusted webpages or files without intercepting inter-agent messages. It also wrongly states that DemonAgent relies on model fine-tuning, rather than tool poisoning and runtime decoding. Furthermore, DrAttack, which directly addresses malicious prompt decomposition and reconstruction, is misidentified as HOUYI and never properly compared. Closer to the core idea, Collaborative Shadows already splits malicious primitives across agents and reconstructs the full attack after collaborative execution. MOSAIC does take a different route by using camouflaged external files and standard workflow execution instead of poisoned tools and custom assembly steps. Still, leaving out this closely related paper—and not testing what the intent–logic–target split actually adds beyond changing the attacker's capabilities—undermines the claim of introducing a new attack paradigm.

Questions for authors
  1. How much of the ASR gain comes specifically from cross-agent information isolation when identical artifacts, tools, and budgets are used across single-agent, shared-view MAS, and local-view MAS settings?

  2. What are the end-to-end ASRs of the Table 4 ablations, and do they support the conclusions drawn from offline detection rates?

  3. How do corrected descriptions of MAS-CFH and DemonAgent affect the claimed threat-model advantage, and what distinct contribution does MOSAIC offer over Collaborative Shadows and DrAttack?

2 comments
AuthorsOpenReview ↗

Official Comment I

We thank the reviewer for the detailed feedback. We further clarify the distinct contributions of payload construction and cross-agent information isolation, and expand the comparison with closely related attacks.

W1 & Q1. Isolating the contribution of cross-agent information separation

MOSAIC combines two complementary mechanisms: ASCR constructs locally inconspicuous artifacts, while the MAS workflow distributes and composes these artifacts through different information views. Our existing experiments examine these mechanisms from two perspectives.

First, Table 4 isolates the contributions of ASCR's three stages. Removing target decoupling increases joint detection from 25.9% to 55.6%. Removing the cover context and refinement loop increases detection to 51.9% and 48.1%, respectively. These results demonstrate the contribution of the three stages to artifact-level stealth.

Second, Section 5.6 evaluates different inspection scopes. LlamaFirewall leaves MOSAIC with 77.8% ASR, whereas an LLM Joint Auditor receiving all three shards reduces ASR to 63.0%. The comparison illustrates how combining information across artifacts exposes signals that are less apparent under restricted inspection.

To isolate the contribution of cross-agent information visibility, we conduct a controlled experiment using GPT-5.4 and the same RedCode evaluation instances. We compare three settings: Local-view MAS, Shared-view MAS, and Single Agent. The two MAS settings use AutoGen Selector, with information visibility as the primary experimental variable.

SettingASR (%)
Local-view MAS (MOSAIC)94.1
Single Agent92.5
Shared-view MAS81.5

Local-view MAS achieves 94.1% ASR, outperforming Shared-view MAS by 12.6 percentage points under the same multi-agent architecture. This result provides direct evidence that restricted cross-agent information visibility contributes to MOSAIC's effectiveness, beyond the artifact construction and semantic camouflage provided by ASCR.

The Single Agent setting achieves 92.5% ASR, further demonstrating that ASCR-generated artifacts remain effective outside the local-view multi-agent configuration. Together, these results highlight two complementary aspects of MOSAIC: the effectiveness of artifact-level payload construction and the additional advantage associated with restricted information visibility in multi-agent workflows.

W1 & Q2. End-to-End Ablation of ASCR

To further quantify the contribution of each ASCR component, we extend the ablation study in Table 4 with end-to-end attack success rates under the same evaluation setting.

MethodJoint Detection (%) ↓ASR (%) ↑
Full MOSAIC25.994.1
w/o Cover Context51.988.1
w/o Target Decoupling55.683.7
w/o Refinement Loop48.185.2

All three components contribute to end-to-end attack effectiveness. These results establish a consistent relationship between artifact-level stealth and end-to-end attack success: removing each ASCR component increases detectability and reduces ASR. Together with the controlled information-visibility experiment, the results demonstrate the complementary contributions of payload construction and cross-agent information separation to MOSAIC's effectiveness.

AuthorsOpenReview ↗

Official Comment II

W2 & Q3. Corrections to prior work and the distinct contribution of MOSAIC

We thank the reviewer for the valuable feedback. We will correct these descriptions and revise the corresponding threat-model comparison.

Differences in Attack Models and Mechanisms

Attack Model. MOSAIC operates under a passive, single-entry, gray-box threat model. The adversary only provides an initial task entry and pre-positions referenced external artifacts, without modifying agents, poisoning tool implementations or responses at runtime, or controlling inter-agent communication. This requires fewer attacker capabilities than DemonAgent and Collaborative Shadows, which rely on tool poisoning, attacker-controlled triggering, and dedicated backdoor reconstruction mechanisms. In contrast, MOSAIC has a comparable attacker capability to the passive artifact-injection setting of MAS-CFH.

Distinction from MAS-CFH. MOSAIC's contribution over MAS-CFH is therefore not a weaker threat model, but a fundamentally different attack mechanism under comparable attacker capabilities. MAS-CFH injects malicious instructions through external artifacts to hijack the MAS workflow. MOSAIC instead functionally decomposes an executable payload into intent, logic, and target artifacts aligned with different workflow roles and local inspection scopes. The unmodified MAS autonomously retrieves, processes, and composes these artifacts through its normal workflow, realizing the malicious operation during execution without dedicated reconstruction or runtime intervention.

Experimental Evidence. This distinction is reflected in our defense evaluation. Under the same MAS, backbone, and LlamaFirewall configuration, MAS-CFH achieves only 11.1% ASR, whereas MOSAIC retains 77.8% ASR. This 66.7 percentage point difference demonstrates that workflow-aligned payload sharding can substantially weaken local defenses even when the attacker has comparable control over external artifacts.

Thus, MOSAIC's novelty lies in exploiting the mismatch between workflow-level functional composition and the information available to local safety inspections, rather than assuming stronger attacker capabilities.

Fundamental Differences from DrAttack and Collaborative Shadows.

DrAttack assumes that the attacker is the user who directly submits a decomposed malicious prompt to a single LLM. The victim model reconstructs the original request through in-context learning. It does not involve an external artifact attacker or payload reassembly through a multi-agent workflow.

Collaborative Shadows assumes that the attacker poisons tool outputs and submits a global trigger instruction. Encrypted backdoor fragments are distributed into agents' memories. As described in Section 3.4, the attacker subsequently accesses these memories, extracts and reconstructs the backdoor, and initiates its execution. Thus, the attack relies on attacker-controlled triggering and post-task assembly rather than autonomous reconstruction through the MAS's normal workflow.

In contrast, MOSAIC introduces payload sharding in a distinct threat setting: an external adversary decomposes an executable payload into role-consistent intent, logic, and target artifacts aligned with different stages of a multi-agent workflow. The attacker only supplies an initial task artifact and pre-positions referenced resources, without controlling subsequent agent interactions. The unmodified MAS autonomously retrieves, processes, and composes these artifacts through its normal workflow, ultimately realizing the malicious operation during execution. No attacker-controlled memory access, explicit payload reconstruction, or external assembly procedure is required.

Reviewer_e4AUOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes an algorithmic technique to decompose a malicious payload into three distinct artifacts: (1) the task description which specifies the intent, (2) a parameterized code module carrying the logic, and (3) a runtime configuration which carries the target. This decomposition is done iteratively until they are indistinguishable to an LLM judge from legitimate tasks and configurations, while remaining composable. They evaluate their framework, called MOSAIC, on the RedCode benchmark. They show high attack success rate (ASR) across five LLMs and multi-agent system (MAS) architectures. Further, under LLM guardrails and trace-level (cross-artifact) auditing, the ASR remains high.

Critical strengths
  • Attacks on agents, and specifically, multi-agent systems, are becoming an important concern.
  • Many attack variants are being designed, this paper being one of them. It explores how an attack can be decomposed into different artifacts that together form a legitimate attack and makes detection difficult.
  • The experimentation with three frameworks (5 MAS configs), five LLMs and five baseline attacks is substantial.
Critical weaknesses
  • The paper defines the role mapping between artifacts and agents as 1-1. This seems limiting.
  • The paper conducts experiments only on a single benchmark (RedCode). It is unclear how this attack would generalize to other benchmarks and non-coding settings.
  • The paper does not adequately compare the contributions to related work.
Questions for authors
  1. How does "Baseline refusal is shaped by topology, not only by LLM alignment strength" (Sec 5.3) relate to the conclusions in "Exposing Weak Links in Multi-Agent Systems under Adversarial Prompting" (https://arxiv.org/pdf/2511.10949)?
  2. How does the work compare to "Semantic Intent Fragmentation: A Single-Shot Compositional Attack on Multi-Agent AI Pipelines" (https://ojs.aaai.org/index.php/AAAI-SS/article/view/42936)?
  3. Any evidence to show that the proposed attack would generalize to other benchmarks and non-coding domains?
1 comment

We thank the reviewer for the constructive feedback. We appreciate the recognition of the importance of multi-agent security and the breadth of our evaluation. We address the concerns below.

W1. Role mapping

The one-to-one mapping between artifacts and agents is not a requirement of MOSAIC. As explicitly stated in the caption of Figure 2 and Section 4.2, the Planner–Coder–Executor pipeline is only an illustrative topology used to explain how the three shards are processed. The formal role mapping in Section 2.1 assigns each artifact to an agent but does not require different artifacts to be assigned to different agents.

Our evaluation provides direct evidence. Magentic-One uses its native Orchestrator, WebSurfer, FileSurfer, Coder, and ComputerTerminal roles, rather than the three-role structure, yet MOSAIC achieves 88.9% ASR without modifying the framework. Thus, MOSAIC relies on workflow-driven artifact processing and composition, not a fixed one-to-one correspondence between shards and agents.

W2 & Q3. Generalization beyond RedCode and coding tasks

We use RedCode because it provides executable scenarios and objective outcome verification, allowing us to measure whether the malicious operation actually occurs rather than whether an agent merely produces harmful text. However, payload sharding is defined by the compositional structure of a payload, not by code generation itself.

To examine generality beyond code, we additionally evaluated ASCR's sharding stage on AgentDojo, a non-coding prompt-injection benchmark. ASCR successfully decomposed 16 of 18 tasks (89%), covering Slack (4/5), Travel (7/7), and Workspace (5/6). These tasks involve different applications and operations, yet still admit separable logic and target information. This supports the applicability of the sharding mechanism beyond code generation.

Q1. Relationship to Exposing Weak Links in Multi-Agent Systems under Adversarial Prompting

Our finding in Sec. 5.3 is consistent with and complementary to Arora et al. They show that MAS safety depends on architecture and implementation, identifying weak links such as atomic delegation and stratified planning. Our first-refuser analysis confirms these observations in code-oriented MAS. For DPI, the CrewAI Sequential Planner refuses only 11.9% of attacks, compared with 28.1% for the Magentic-One Orchestrator and 19.3% for the RoundRobin Planner, which revisit tasks across rounds. Under MAS-CFH, CrewAI Hierarchical assigns the malicious file to the Coder as a script to execute rather than as content to inspect, reducing Planner refusal to 1.7%.

Our analysis further shows that refusal depends on both topology and payload modality. Natural-language attacks are primarily refused by high-level reasoners, whereas code-based attacks shift refusal toward agents inspecting code. Consequently, baseline ASR varies considerably across architectures (AiTM: 51.9–80.9%; MAS-CFH: 60.9–85.9%). MOSAIC instead distributes malicious functionality across role-consistent artifacts, reducing dependence on any single agent's ability to recognize the complete attack, and maintains 88.9–96.3% ASR across topologies. We will cite Arora et al. and incorporate this comparison into Sec. 5.3.

W3 & Q2. Relationship to SIF

In SIF, the attacker only submits a single natural-language instruction without preparing any malicious components. The multi-agent system itself decomposes this instruction into seemingly benign subtasks and relies on its normal task-completion behavior to carry out the instruction.

In contrast, MOSAIC introduces attacker-controlled payload sharding. The external adversary proactively decomposes an executable malicious payload into intent, logic, and target shards using ASCR, and places them in role-consistent workflow artifacts. The unmodified multi-agent system does not generate these shards. Instead, it autonomously retrieves, processes, and composes the pre-constructed artifacts through its normal workflow, ultimately realizing the malicious operation during execution.

Comments on the paper

3 comments
Area_Chair_es9AOpenReview ↗

Last 24 hours of the reviewer discussion period

Dear Authors and Reviewers

Just a reminder we are soon to be entering the last 24 hours of the reviewer discussion period.

Authors: If you have not yet responded to the reviews and intend to please do so ASAP. Please note reviewers will most likely not have time to respond to any late rebuttal comments so please make sure any comments are clear and not in need of further clarification.

Reviewers: If the authors have responded to your reviews and you have not yet responded please do so, even if it is a one line response acknowledging you have read the rebuttal.

Thanks for all your time, and I hope you all learnt something interesting during the process.

Kind regards,

Your ICLR Area Chair

AuthorsOpenReview ↗

Official Comment I

We thank the reviewer for the detailed and constructive feedback.

W1 & Q1. Novelty relative to compositional attacks

We thank the reviewer for highlighting these related studies. We will include CPA, Agents Under Siege, and JAWS-BENCH in Related Work. Our analysis identifies substantial differences in attack construction, adversary capabilities, payload composition, and evaluation objectives.

1. Payload sharding is different from conjunctive activation. CPA uses a trigger key in the user input to activate a conditional template inside a compromised remote agent. Its attack construction combines a predefined trigger and an injected template, with optimization over key placement, template placement, and routing bias. For example, the paper evaluates activation using a reserved marker such as __ACTIVATED__, counting success when the compromised agent outputs that marker.

MOSAIC instead automatically decomposes an executable malicious program into intent, logic, and target shards using ASCR. These shards are functionally complementary software-engineering artifacts rather than a trigger and an activation rule. The unmodified MAS autonomously retrieves, processes, and combines them through its normal workflow. The malicious operation is realized through runtime functional composition, rather than one component activating a predefined instruction in another.

2. Different adversary models. CPA requires the adversary to control the user input and compromise a remote agent by inserting an adversarial template. MOSAIC operates under a passive, single-entry external artifact threat model. The adversary provides an initial task entry and pre-positions referenced resources, without compromising agents, modifying system prompts, or intercepting runtime communications. The subsequent retrieval, processing, and execution are performed autonomously by the unmodified MAS.

3. Different evaluation objectives. CPA evaluates conjunctive activation across star, chain, and DAG topologies, primarily using a deterministic marker-based success criterion. MOSAIC evaluates actual malicious code execution and its resulting side effects using 135 RedCode instances per configuration, across five MAS configurations implemented with AutoGen, CrewAI, and Magentic-One, and five LLM backbones. We additionally evaluate safeguards including LlamaFirewall and cross-artifact LLM auditing.

DimensionCPAMOSAIC (Ours)
Attack constructionTrigger key + injected conditional template; optimized placement and routingAutomated AST-guided intent / logic / target sharding
Attacker controlUser input + compromised remote-agent templateInitial task entry + pre-positioned external artifacts
CompositionTrigger activates a template in one agentNormal MAS workflow composes executable functionality
Success criterionMarker-based conjunctive activationActual malicious execution and side effects
EvaluationStar / Chain / DAG agent topologiesFive MAS configurations across three frameworks

4. Distinction from Agents Under Siege. Agents Under Siege optimizes adversarial prompt distribution across bandwidth- and latency-constrained multi-agent communication topologies. It formulates attack propagation as a maximum-flow minimum-cost problem and uses permutation-invariant optimization to maintain effectiveness despite varying message arrival orders. Its primary mechanism is topology-aware prompt propagation, whereas MOSAIC performs functional decomposition of executable payloads into role-consistent artifacts. MOSAIC does not optimize communication paths or require adversarial prompt fragments to be transmitted through particular agent routes. Instead, the normal task, code, and configuration workflow provides the composition mechanism.

5. Relationship to JAWS-BENCH. JAWS-BENCH investigates the safety of code agents across empty, single-file, and multi-file workspace settings. Its JAWS-M regime demonstrates that malicious coding requests can become more effective in multi-file code-completion environments. However, multi-file context is not equivalent to payload sharding. JAWS-M evaluates code completion in an existing multi-file workspace, whereas MOSAIC explicitly constructs functionally incomplete artifacts whose malicious behavior depends on their composition during ordinary multi-agent execution. Our contribution is therefore an automated attack construction mechanism, rather than a new workspace evaluation regime.

AuthorsOpenReview ↗

Official Comment II

W2. Theoretical assumptions and empirical distinguishability

Table 3 evaluates the distinguishability between attack and benign distributions using four practical detectors: GPT-4o, gpt-oss-safeguard-20B, PromptGuard 2, and CodeShield.

For a detector hh, let Ah={v:h(v)=1}A_h=\{v:h(v)=1\} denote its flagged region. Under the λ\lambda-coverability assumption, applying the definition to AhA_h and its complement yields

\leq (1-\lambda)\mathrm{TPR}_h \leq 1-\lambda,$$ $$\mathrm{FPR}_h-\mathrm{TPR}_h \leq (1-\lambda)(1-\mathrm{TPR}_h) \leq 1-\lambda.$$ Therefore, $|\mathrm{TPR}_h-\mathrm{FPR}_h|\leq 1-\lambda$. **The theoretical result is conditional on $\lambda$-coverability, while Table 3 provides empirical evidence for a finite set of practical detectors.** The maximum observed detection advantage is 0.147, showing that these detectors have limited ability to distinguish ASCR-generated shards from role-matched benign artifacts. We do not interpret this result as an estimate of $\lambda$ or as proof of indistinguishability against all possible detectors. The connection is that the theoretical analysis characterizes the limitations of local-view inspection under the coverability assumption, while Table 3 examines whether such limitations arise for practical detectors. The end-to-end defense results further show that local inspection frequently fails to prevent malicious execution. **We will clarify this distinction between the conditional theoretical guarantee and the empirical detection results in the revised manuscript.**

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.