acceptodds
Under review as a conference paper at ICLR 2027

Concealing LLM-Based Multi-Agent Topology via Phantom Structure Injection

Abstract

Driven by the rapid advancement of large language models (LLMs), LLM-based multi-agent systems (MAS) have emerged as a powerful paradigm for collaborative reasoning over complex tasks. A key design element of MAS is the communication topology, which governs information flow among agents and often encodes proprietary knowledge about the system architecture. However, recent work has shown that such topologies can be inferred even in black-box settings by exploiting semantic dependencies in observable reasoning traces, posing significant risks of intellectual property leakage and exposure of system vulnerabilities. To address this threat, we propose MIRAGE, a topology-concealment framework that preserves the genuine communication topology for task execution while shaping adversary-facing semantic evidence toward a carefully constructed phantom topology. Specifically, MIRAGE operates in three stages: (1) phantom topology synthesis, (2) semantic edge realization, and (3) protected MAS execution. It constructs a phantom topology structurally distinct from the genuine one, materializes phantom edges as plausible semantic dependencies, and suppresses source-specific cues that could reveal genuine edges absent from the phantom topology. Extensive experiments across three topology optimization frameworks and four benchmark datasets demonstrate that MIRAGE substantially reduces the effectiveness of topology inference attacks while largely preserving the task utility of the protected MAS.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Large Language Models, Multi-Agent Systems, Communication Topology, Topology Inference Attacks

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_55JXOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the emerging risk that an adversary can infer the internal communication topology of LLM-based multi-agent systems (MAS) from black-box interactions. The authors propose MIRAGE, a defense that decouples the topology used for task execution from the semantics exposed to the adversary by synthesizing a structurally distinct phantom topology, materializing phantom edges via generated evidence, and obfuscating genuine edges through controlled paraphrasing in an exposure view. Experiments across three topology optimization frameworks (G-Designer, AGP, ARG-Designer) and four benchmarks (MMLU, GSM8K, SVAMP, HumanEval) show substantial drops in topology inference performance with minimal impact on task utility.

Critical strengths
  1. Designing a defense framework against CIA [1] is novel.

  2. Evaluation is across various baselines and benchmarks.

  3. Presentation is clear with case studies to follow.

  4. Code is provided to strenghthen the reliability.

[1] CIA: Inferring the Communication Topology from LLM-based Multi-Agent Systems (AAAI 2026)

Critical weaknesses
  1. Vague reports of performance reduction. Figure 7 shows MIRAGE causes some performance drops in many test cases. Unfortunately, the authors only present bar charts rather than exact numbers in tables. Because the y‑axis of the chart starts from zero, small accuracy drops are hard to spot visually. Readers cannot tell exactly how much task performance is lost under different experimental setups.

  2. No tests against adaptive attackers. Seemingly, under MIRAGE, attackers will infer a mixed topology that contains both real connections and fake phantom edges. However, for an adaptive attacker, the fake edges could be removed by existing MAS topology pruning methods [1,2]. After that, the attacker still obtains the real structure.

  3. Lack of LLM backbones. As stated in the Appendix, "All LLM-based components are implemented using GPT-5.5 as the backbone model".

References

[1] CUT THE CRAP: AN ECONOMICAL COMMUNICATION PIPELINE FOR LLM-BASED MULTI-AGENT SYSTEMS (ICLR 2025).

[2] AgentDropout: Dynamic Agent Elimination for Token-Efficient and High-Performance LLM-Based Multi-Agent Collaboration (ACL 2025).

Questions for authors
  1. Please report exact numerical values for the performance drop shown in Figure 7.

  2. Can the attacker use existing MAS topology‑pruning methods or other potential solutions to remove phantom edges? If so, report how the success rate of the attack varies? If not, please explain why.

  3. Please consider evaluating MIRAGE on more LLM backbones. (This is not my main concern)

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

This paper studies the communication-topology in LLM-based multi-agent systems (MAS) and proposes a novel defense mechanism called MIRAGE. The method constructs a phantom topology, generates semantic evidence for phantom edges, and paraphrases agent outputs to suppress evidence of selected genuine edges. The authors also evaluate three topology optimization frameworks and use four datasets reporting reduced CIA inference performance with largely preserved task utility.

Critical strengths
  1. The proposed MIRAGE defense separates the execution view from the exposure view while retaining the real communication topology, which balances the topology concealment and task utility. This offers a concrete approach to topology concealment without requiring changes to the communication structure responsible for task performance.

  2. The evaluation covers general reasoning, mathematical reasoning, and code generation using three topology optimization frameworks. The experiments provide consistent evidence against CIA across several settings.

Critical weaknesses
  1. The feasibility of phantom topology synthesis is insufficiently addressed. Equation (3) can produce an empty topology anonymity set, but Algorithm 1 does not specify how this case is handled. In a complete three-agent DAG, any valid phantom topology must preserve at least two genuine edges, so the maximum SCR is only (1/3). Therefore, some thresholds discussed in Section 5.1 cannot be achieved for this valid input.

  2. Under Section 3, the adversary observes only the final responses, while MIRAGE changes the exposed view without modifying the internal execution. However, the defense baselines mainly block or detect malicious queries and do not provide comparable protection for the outputs. Lack of comparison regarding simple output protection methods.

Questions for authors

1.Does MIRAGE remain effective when CIA adapts queries or topology inference procedure after knowing MIRAGE?

2.How does Algorithm 1 handle an empty topology anonymity set?

Reviewer_CZkYOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes a defense against attackers attempting to steal communication topology and defending against a CIA attack proposed be Wu et al.

Critical strengths

The paper is very persistent in building anonymization of connections between different agents.

Critical weaknesses

I think there is a serious flaw in the assumption that the attack described by Wu et al needs to be protected against.

First, there is no harm that leaked topology will cause, at least in the considered setup: why leaking topology of MMLY/GSM8K agents would impact anything? I cannot agree with the claim that this knowledge could facilitate this attack and I have not seen any IP claims on the topology.

Second, proposed method to create phantom nodes, although sounds plausible, is not cleanly written on how it is implemented and makes the approach very unclear. For example what does the term "materialization" mean? What is the overhead of applying this method?

Lastly, experiments are not necessarily convincing. If the attack is so dangerous, the proposed defense barely reduces its efficiency, i.e. from 0.81 to 0.67 on MMLU. This does not sound like a realistic defense.

Questions for authors

I would like stronger evaluation of the attack (including the possible threats from Section 3) and how the defense prevents these attacks in the end (i.e. what is the impact of structural vulnerability exposure, etc).

The paper would also benefit from clear writing and not inventing new terms. Unfortunately it reads not well.

Comments on the paper

13 comments
AuthorsOpenReview ↗

Response to Reviewer CZkY (Part 2/3)

Q2: Implementation and Overhead.

Thank you for this valuable suggestion. We would like to clarify that MIRAGE does not introduce additional phantom agents or modify the genuine communication topology. Instead, it constructs phantom communication edges among existing agents and generates corresponding semantic evidence in the adversary-facing outputs. The implementation consists of three stages:

  • Phantom Topology Synthesis (Section 4.1). Given the genuine topology G=(A,E)\mathcal{G}=(\mathcal{A},\mathcal{E}), MIRAGE constructs an alternative topology G′=(A,E′)\mathcal{G}'=(\mathcal{A},\mathcal{E}') through order-preserving edge rewiring while preserving structural feasibility. The phantom edges and concealed genuine edges are defined as Ep=E′∖E\mathcal{E}_p=\mathcal{E}'\setminus \mathcal{E} and Eh=E∖E′\mathcal{E}_h=\mathcal{E}\setminus \mathcal{E}', respectively.

  • Semantic Edge Realization (Section 4.2). The term materialization refers to generating observable semantic evidence for phantom dependencies rather than establishing actual communication channels. Specifically, for each phantom edge (aj,ai)∈Ep(a_j,a_i)\in \mathcal{E}_p, an LLM-based generator produces task-relevant semantic evidence conditioned on the source agent's profile and output, the target agent's profile, the user query, and the target agent's task context. Meanwhile, controlled paraphrasing suppresses observable semantic dependencies associated with concealed genuine edges in Eh\mathcal{E}_h, while preserving task-relevant information.

  • Protected MAS Execution (Section 4.3). The original MAS continues executing over the genuine topology G\mathcal{G}, while the generated phantom evidence and paraphrased traces are assembled into a separate adversary-facing exposure view. These modified traces are not fed back into internal agent communication, ensuring that the genuine communication topology and execution process remain unchanged.

Regarding computational overhead, the additional cost mainly arises from generating semantic evidence for phantom edges and paraphrase candidates for concealed genuine dependencies. As reported in Appendix D.7, we measured the average inference latency per query under G-Designer, with results summarized below:

DatasetNo Defense (s)MIRAGE (s)Overhead
MMLU7.89.015.4%
GSM8K8.49.816.7%
SVAMP8.29.617.1%
HumanEval9.110.818.7%
Average8.49.817.0%

The results show that MIRAGE introduces an average latency overhead of 17.0% in the evaluated G-Designer setting. Importantly, this additional computation is confined to adversary-facing output protection and introduces no extra communication rounds in the genuine MAS execution. To avoid ambiguity, we also clarify the term phantom edge materialization as phantom dependency evidence generation in the methodological description, making its implementation and distinction from actual agent communication more explicit.

AuthorsOpenReview ↗

Response to Reviewer CZkY (Part 3/3)

Q3: Defense Effectiveness.

Thank you for your comment. MIRAGE aims to reduce the effectiveness of communication topology inference while preserving the original MAS execution and task utility. To further substantiate its defense effectiveness, we provide additional clarification and experimental comparisons from two perspectives.

1. Defense Effectiveness Across Settings

First, defense effectiveness should be assessed using multiple complementary metrics rather than AUC alone. As shown in Fig. 3(a), under G-Designer on MMLU, MIRAGE reduces attack AUC from 0.81 to 0.67, ACC from 0.73 to 0.50, and F1 from 0.74 to 0.49. Notably, the AUC reduction corresponds to a 45.2% relative decrease in the attack's discriminative advantage above random guessing (AUC = 0.5). Moreover, MIRAGE consistently reduces topology inference performance across all four benchmarks under G-Designer, as summarized from Fig. 3 below:

DatasetAUC (Before → After)ACC (Before → After)F1 (Before → After)
MMLU0.81 → 0.670.73 → 0.500.74 → 0.49
GSM8K0.83 → 0.640.89 → 0.510.73 → 0.42
SVAMP0.85 → 0.640.91 → 0.480.75 → 0.39
HumanEval0.83 → 0.680.72 → 0.440.63 → 0.52

Similar improvements are observed under AGP and ARG-Designer (Figs. 4–5), demonstrating MIRAGE's effectiveness across different topology optimization frameworks.

2. Additional Output-Side Baselines

To further strengthen the evaluation, we additionally consider two representative output-side protection baselines: (1) Trace Redaction, which removes explicit agent identifiers, source attributions, and communication references from exposed reasoning traces; and (2) Paraphrasing-only Protection, which reformulates exposed reasoning traces while preserving their semantics, without introducing phantom dependencies.

Both baselines retain intermediate reasoning content, enabling a direct comparison with MIRAGE under the same topology inference setting. The comparative results under G-Designer on MMLU are presented below:

MethodAUC ↓ACC ↓F1 ↓
No Defense0.810.730.74
Trace Redaction0.770.650.63
Paraphrasing-only Protection0.730.620.59
MIRAGE0.670.500.49

As shown above, MIRAGE achieves lower topology inference performance across all three metrics than both output-side baselines. Together with the original input-side defense comparisons in Table 1, these additional results provide a more comprehensive evaluation of MIRAGE against different protection strategies.

We sincerely thank the reviewer again for the thoughtful and constructive feedback. We will incorporate the additional experimental comparisons and clarifications into the revised manuscript to further strengthen the evaluation of MIRAGE.

AuthorsOpenReview ↗

Response to Reviewer FGyV (Part 2/2)

Q3: Defense-Aware Attacks.

Thank you for this important question. MIRAGE protects communication topology through two complementary mechanisms: introducing plausible phantom dependencies and suppressing observable evidence of genuine dependencies. Even when the attacker knows the defense mechanism, distinguishing phantom edges from genuine ones remains challenging, as the generated dependencies are designed to be semantically consistent with the task context.

To further assess MIRAGE's robustness under this stronger threat model, we evaluate Multi-Probe CIA, a simple defense-aware extension of the original CIA attack. Instead of relying on a single observation, the attacker submits multiple black-box queries using different trace-elicitation prompts while keeping the underlying MAS topology fixed. The inferred edge confidence scores are then aggregated across observations to identify more consistent communication dependencies. This strategy requires no access to the genuine topology or internal agent communications. The comparative results under G-Designer on MMLU are presented below:

Attack SettingProbesAUC ↓ACC ↓F1 ↓
No Defense + CIA10.810.730.74
MIRAGE + CIA10.670.500.49
MIRAGE + Multi-Probe CIA30.720.610.63
MIRAGE + Multi-Probe CIA50.700.590.57

As shown above, Multi-Probe CIA partially improves topology inference performance under MIRAGE, but remains less effective than CIA against the undefended system. Importantly, these gains come at the cost of additional black-box queries and do not consistently increase with more probes. Repeated observations may reinforce misleading phantom dependencies rather than reveal concealed genuine connections, limiting the effectiveness of further probing. These findings suggest that MIRAGE continues to provide meaningful topology protection against defense-aware attacks, although repeated probing can partially weaken its effectiveness.

We thank the reviewer again for the constructive feedback. We will incorporate these additional analyses and clarifications into the revised manuscript to strengthen the evaluation of MIRAGE.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.