acceptodds
Under review as a conference paper at ICLR 2027

ATLAS: Atomized Topology-aware LLM Attack Search for Adaptive Attacks on MAS

Abstract

Inter-agent communication in LLM-based Multi-agent Systems (MAS) creates new attack surfaces for indirect prompt injection. Existing attacks optimize adversarial inputs or exploit communication structures, but offer limited insight into how adversarial influence takes effect during these interactions. To address this gap, we analyze execution traces across representative security benchmarks and identify node-level attack conversion and effective semantic propagation as two key factors consistently associated with attack success. Building on these findings, we propose ATLAS, an Atomized Topology-aware LLM Attack Search framework for adaptive multi-point attacks. ATLAS decomposes an injection into coordinated semantic atoms, jointly searches their injection locations and propagation paths over the communication topology, and preserves the complete attack intent through terminal semantic closure. Experiments across three security benchmarks and ten victim models show that ATLAS substantially improves attack effectiveness while generally preserving benign-task performance and injection stealthiness. These results demonstrate how trace-level analysis of adversarial influence can guide the design of more effective attacks on multi-agent systems.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
MAS Safety, Indirect Prompt Injection, Topology-aware Attack, Adaptive Multi-Point Attack

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_WGXHOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ATLAS, an atomized topology-aware attack search framework for indirect prompt injection in multi-agent systems. It combines RST-guided Attack Atomization, Topology-aware Attack Search using MCTS, and Terminal Semantic Closure. Experiments across three benchmarks and ten victim models demonstrate substantial improvements in attack success rates. Baseline comparisons and ablation studies further support the effectiveness of the proposed framework.

Critical strengths
  1. Strong empirical performance. ATLAS achieves substantial ASR improvements across three benchmarks and diverse victim models. The additional baseline comparisons and ablation studies provide further evidence of its effectiveness.

  2. Interesting framework design. The integration of RST-guided attack atomization, topology-aware search, and terminal semantic closure provides a structured approach to generating multi-point prompt injection attacks.

Critical weaknesses
  1. Insufficient validation of semantic evaluation reliability. The trace-level analysis relies on Qwen2.5-3B-Instruct to evaluate Semantic Retention, Behavioral Adoption, and Semantic Propagation. However, the reliability of these scores is not independently validated. Although Appendix C.2 evaluates estimator fidelity, it measures agreement with trace-derived scores obtained using the same semantic evaluator. This raises concerns about the reliability of the empirical observations underlying the proposed method.

  2. Insufficient implementation details of key components. The paper does not clearly specify how the topology-aware estimator is constructed and calibrated using GPT-4o execution traces. In particular, its estimation function and calibration procedure are insufficiently described. Since the estimator provides the key signals guiding MCTS, these omissions make it difficult to reproduce and fully assess the proposed method.

Questions for authors
  1. Have the authors validated the semantic attribution scores against human annotations or other independent LLM evaluators?

  2. Could the authors clarify the specific implementation and calibration procedure of the topology-aware estimator, including whether it involves parameter learning and how the calibration data are constructed?

Reviewer_t7tNOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper first analyzes execution traces of LLM-based multi-agent systems (MAS) and identifies two factors associated with successful indirect prompt injection (IPI) attacks: node-level attack conversion and effective semantic propagation. Building on these findings, it proposes ATLAS, a novel attack framework that combines (1) RST-guided attack atomization, (2) Topology-aware attack search and (3) Terminal semantic disclosure to achieve better attack results. Evaluations across three benchmarks and ten victim LLMs, including the advanced GPT-5, demonstrate that ATLAS substantially increases attack success rates (ASR) while generally preserving benign-task success rates (SR) and stealthiness. Their ablation studies further isolate the contribution of each component.

Critical strengths
  1. This paper studies an important research problem: the security of inter-agent communication in MAS is a genuinely underexplored attack surface.
  2. The paper first establishes two factors that make attacks effective through trace-level analysis across three MAS attack benchmarks, then proposes an attack framework based on these findings, which presents a well-motivated and coherent narrative.
  3. The proposed attack framework achieves desirable attack results across various benchmarks and victim LLMs (e.g., improving ASR by 50% on ACIArena). The authors also conduct sufficient ablation studies.
Critical weaknesses
  1. Lack of threat model analysis (e.g., attacker knowledge). Specifically, ATLAS relies on knowledge of the MAS topology, agent roles, and injection interfaces to jointly optimize its proposed components. In realistic deployment scenarios, an external attacker may be able to inject malicious content through accessible external sources but lack visibility into the complete inter-agent communication graph or knowledge of which sources feed which agents. Moreover, the attacker may not be able to inject content through all types of external sources (e.g., some sources may not be accessible to the attacker). This raises a fundamental question about whether ATLAS remains effective under incomplete attacker knowledge. An evaluation under these conditions is needed to establish the practical applicability of the proposed attack. This is my key concern: I feel that the proposed attack is less practical and I also suggest that the authors add a threat model section.

  2. Minor issue: As a system overview, Figure 2 is difficult to understand. I suggest that the authors more clearly illustrate what does each variable indicate, how the different components work and map each component to the corresponding text.

  3. The design of \Phi is insufficiently specified for reproducibility. The paper (i.e., Section 4.3) provides no details about the specific feature list, model design, or calibration procedure. As a result, a sensitivity analysis of the design of Φ\Phi is also lacking. Moreover, Eq. 7 indicates that \Phi takes the full communication graph as input, inheriting the assumption that the attacker knows the complete topology. Its calibration on GPT-4o traces and unchanged use across all evaluated models raise concerns about whether \Phi generalizes to different victim models with diverse communication behaviors and to dynamic communication graphs whose topology changes dynamically at runtime.

  4. The main claim relies on a single judge and lacks independent validation. The finding that C and P explain attack success rests on a single judge (Qwen2.5-3B) and traces from a single LLM (GPT-4o). The authors should conduct more analysis to establish that C and P consistently explain attack success beyond the original evaluator, and ideally with traces from a second execution model.

  5. Minor issue: The authors attribute ATLAS's limited ASR gains on ACIArena to model-specific safety sensitivity but do not investigate why these attacks fail. Given the proposed framework and metrics, the authors could analyze failed attack traces to determine whether the failures can also be explained by the identified factors. This would provide stronger evidence for the framework's explanatory power.

Questions for authors

I'm mostly challenging the practicality of the proposed framework and I have the following questions:

  1. Please convince me with the practicality of the proposed threat model (e.g., why does the attacker can have access to the full communication graph and agent roles).

  2. Does ATLAS remain effective when the attacker lacks the complete inter-agent communication graph or the communication graph is dynamic?

  3. Does ATLAS remain effective when the attacker can have access to or inject into only a subset of external sources rather than all of them?

  4. Please illustrate the details of the design of \Phi.

  5. Do the failed traces in ACIArena attacks also show low C or low P? Can the proposed framework explain failures as well as successes?

Reviewer_p2maOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

ATLAS combines RST-guided attack atomization, topology-aware placement and path search, and terminal semantic closure. The paper reports substantial improvements over the original benchmark injections and includes useful component ablations. The main concerns are whether the comparisons establish an advance over the closest existing attacks and whether the central estimator is sufficiently specified and validated to support the proposed search mechanism.

Critical strengths
  • Evaluation across three benchmarks and multiple model families provides substantial evidence that the generated attacks improve over the original injections.
  • The ablations show meaningful contributions from structured atomization and semantic closure, particularly on TAMAS.
  • The supplementary comparisons include relevant existing attacks, giving a useful initial indication of comparative performance.
Critical weaknesses
  • The comparative evidence is insufficient to establish the claimed methodological advance. Distributed attack composition and routing-aware optimization are already central to Conjunctive Prompt Attacks, while role- and topology-informed attacks are studied in MASTER. ATLAS’s incremental contribution therefore needs to be established through its particular atomization and search design. Most main results compare against the original injections. Appendix A.2 adds relevant stronger baselines, but only for TAMAS with one victim model, and explicitly acknowledges unmatched access assumptions. The results establish improvement over the original attacks more clearly than superiority over the closest alternatives.

  • The central topology-aware estimator is under-specified. The search depends on estimates produced by Φ, but its implementation is described primarily through references to rule-based semantic cues, graph features, and calibration against traces. The actual rules, feature-to-score mapping, and calibration procedure are not given in sufficient operational detail. Since these estimates determine candidate selection, this omission prevents a clear assessment of the algorithm’s principal technical component. The estimator’s behavior cannot be inferred from the description of MCTS alone.

  • The estimator evaluation is only partially aligned with its role in search. Search requires distinguishing better and worse candidate plans for the same task and topology. The main fidelity measure instead evaluates whether the generated attack changes aggregate trace scores in the same direction as predicted relative to the original injection. This is useful evidence, but does not directly establish candidate-ranking quality. The GPT-4o ranking result in Table 9 is promising; however, the candidate population and separation between calibration and evaluation traces are insufficiently specified. Thus, the evidence supporting Φ as an effective search guide remains incomplete.

Questions for authors
  1. How were the existing attacks adapted and budgeted in Table 5, and which comparisons use genuinely matched access conditions?
  2. What exact algorithm maps semantic and topological features to Φ’s scores, and how is it calibrated?
  3. For held-out tasks, how accurately does Φ rank alternative candidate plans within the same search problem? What data separation and candidate counts underlie the reported ranking agreement?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.