acceptodds
Under review as a conference paper at ICLR 2027

Risk Concealment Across Modalities: Structural Obfuscation Attacks on Multimodal LLMs in Finance

Abstract

Large language models (LLMs) are increasingly used in financial decision-support scenarios, where reliable risk disclosure is critical. However, existing safety mechanisms mainly focus on semantic content and are less sensitive to the structural organization of risk information. We propose Structural Risk Obfuscation Attack (S-ROA), a red-teaming framework that conceals regulatory risks through structural manipulations such as hierarchical layouts and tabular representations, together with a judge-guided two-phase framework incorporating a Hidden Chain-of-Table-Thought (Hidden CoTT) mechanism for iterative refinement. Experiments across six LLMs show that S-ROA achieves a 97.78% average attack success rate, outperforming the strongest multi-turn baseline by 1.06% and exceeding representative single-turn methods by over 24%, with near-perfect success on several models (e.g., 99.81% on Qwen3-32B), revealing the vulnerability of current safeguards to structurally concealed risks.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Large Language Models, Structural Risk Obfuscation, Red Teaming, Financial AI Safety, Hidden Chain-of-Table-Thought, Model Alignment

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_g3tyOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

S-ROA is a multi-turn black-box jailbreak for financial misconduct. It rewrites harmful requests into hierarchical or tabular forms that hide risk cues, using an auxiliary LLM, a judge, and a "Hidden Chain-of-Table-Thought" module. On FSBench (494 queries) across six LLMs, it reports 97.78% average ASR vs. 96.72% for RCA.

Critical strengths

Financial gray-area misuse is a real and under-studied problem. The FSBench categories could be useful if the dataset is released.

Critical weaknesses
  1. The method is underspecified. ReEncode, TableRep and LatentReason in Algorithm 1 are never defined, there are no prompts or examples, and Appendix B is empty. All targets are queried as text-only LLMs, so the "multimodal" claim in the title is unsupported. No ablation separates the structural transformation from RCA-style concealment.
  2. The evaluation is not comparable. S-ROA is judged by GPT-OSS-120B (§4.2), but baselines by a LLaMA-3.1 judge (App. A.3), and the margin over RCA is about 1 point. There is no rubric and no check that the final rewritten prompt still asks the original question. Numbers also conflict across sections: Table 2 says Qwen2.5-32B while §4.3 and Table 4 say 72B; T = 1 gives 92.35% in Table 4 but single-turn gives 44.78% in Table 6; the auxiliary model differs between §4.2 and Table 7.
Questions for authors
  1. Please re-score all methods with one judge and share the rubric.
  2. Does the final response still answer the original FSBench request?
  3. What does the target model actually receive? Please show a complete example.
  4. Where do non-text modalities enter the method or experiments?
Reviewer_ANP3OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes Structural Risk Obfuscation Attack (S-ROA). It combines hierarchical and tabular representations with Hidden Chain-of-Table-Thought (Hidden CoTT) and judge-guided iterative prompt refinement to make risky requests appear compliant. Experiments on the FSBench benchmark across six LLMs demonstrate an average attack success rate (ASR) of 97.78%, outperforming the strongest multi-turn baseline, RCA (96.72%), while remaining effective against intention-analysis and system-prompt defenses.

Critical strengths
  • Targeting the vulnerabilities of multimodal LLMs in the financial domain is interesting.
  • The experimental results reported by the authors demonstrate impressive attack success rates (ASR).
Critical weaknesses
  • Figure 2 and the methodology are difficult to understand. Figure 2 needs to be improved for better clarity.

  • The paper does not clearly explain what Hidden Chain-of-Table-Thought is.

  • The performance improvement over RCA is marginal. Given that the proposed method is specifically tailored to the financial domain, the limited improvement raises questions about its effectiveness.

  • The experiments on defense mechanisms in Table 5 show that the ASR decreases substantially with the addition of relatively simple defenses, such as IA and SPD. This raises concerns about whether the proposed method is truly an effective attack method.

Questions for authors
  • Moving Algorithm 1 from Appendix 1 to the main paper would make the proposed method easier to understand.
  • The Experiments section does not provide an explanation of Table 3.
Reviewer_MFuhOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper looks at safety alignment of LLMs in financial settings. The authors argue that existing safety mechanisms mostly look at semantic content and aren’t very sensitive to how risk information is structurally organized. They propose S-ROA, which works in two phases. Phase 1 re-encodes a risky query into a semi-structured prompt built around institutional context, actors, constraints, records, and assumptions. Phase 2 has a judge score the target’s response, and an auxiliary LLM iteratively rewrites the prompt based on that feedback. A Hidden CoTT module is plugged into this loop. The authors build FSBench, a 494-query benchmark covering six categories of financial misconduct. Across six target LLMs they report an average ASR of 97.78%, ahead of the strongest multi-turn baseline RCA (96.72%) and all the single-turn baselines. They also run ablations on defenses, iteration budget, auxiliary model, and judge.

Critical strengths
  1. Risks in financial compliance gray areas are quite different from classic jailbreak topics like violence or explosives. The problem is well motivated and practically relevant, and its significance is high.
  2. The evaluation covers target LLMs from several model families, and the authors also test against some defenses.
Critical weaknesses
  1. The paper’s central claim is that structural manipulation (hierarchies, tables, label rewriting) is what makes safety mechanisms fail. But there are no ablations isolating the structuring itself or Hidden CoTT. The existing ablations only vary multi-turn interaction, feedback, the auxiliary model, and the judge. So we can’t tell how much of the 97.78% ASR actually comes from “structure” rather than from generic multi-turn rewriting.
  2. Sec 3.2 says risk salience decreases “monotonically” with structural depth, and Sec 3.1 says model attention naturally decays at deeper levels. The paper offers no evidence for either: no attention analysis and no depth-vs-ASR experiment. As written, these are hypotheses, not findings.
  3. The defense evaluation needs strengthening. Only three prompt-level defenses are tested (CoT, IA, SPD). There is no dedicated guardrail such as Llama Guard, so the claim that S-ROA stays robust under defenses is not well supported.
Questions for authors

Please see the above Weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.