acceptodds
Under review as a conference paper at ICLR 2027

Structural Gaming in Financial Safety Alignment: A Dynamic Strategic Evasion Framework via Tabular Representations

Abstract

As large language models (LLMs) are increasingly integrated into financial systems to enforce regulatory guardrails, they fundamentally alter the safety alignment landscape into a co-evolutionary game between platforms and strategic agents. While conventional safety mechanisms are optimized to detect overt semantic risks, they remain largely oblivious to the structural topology of information. This paper formalizes this vulnerability by introducing Structural Gaming into financial safety alignment. We propose Structural Risk Obfuscation Attack (SO-STAG), a red-teaming framework demonstrating how strategic agents can achieve structural compliance arbitrage. SO-STAG systematically camouflages regulatory violations through deep hierarchical layouts and adversarial tabular representations without falsifying the underlying factual content. To operationalize this, we develop a judge-guided two-phase paradigm incorporating a Hidden Chain-of-Table-Thought (HCoTT) mechanism. By exploiting the inherent information asymmetry, SO-STAG forces a pooling equilibrium that traps the defender in an incomplete information state—effectively creating a structural "cognitive fog" where the alignment guardrails can only observe surface tabular representations while remaining blind to the latent malicious reasoning paths. We model this interaction as a dynamic strategic evasion process where the agent iteratively refines its structural representation based on the defender's feedback. Empirical evaluations across six leading LLMs show that SO-STAG achieves a 97.78% average evasion success rate, exposing a pervasive systemic blindness to structurally concealed risks. Crucially, from an algorithmic efficiency perspective, SO-STAG minimizes the agent's computational cost, utilizing over 80% fewer input tokens compared to state-of-the-art semantic multi-turn baselines. Our work highlights a foundational flaw in current alignment paradigms: by failing to account for structural incentives, existing safeguards invite low-cost strategic manipulation, necessitating a shift toward structure-aware mechanism design in financial AI safety.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Large Language Models, Structural Gaming, Financial AI Safety, Structural Risk Obfuscation, Red Teaming, 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_cYS3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies a safety failure mode in financial LLM systems in which harmful intent is obfuscated through hierarchical and tabular representations rather than conventional lexical paraphrasing. It introduces Structural Gaming and proposes SO-STAG, a two-phase, judge-guided red-teaming framework that first re-encodes risky queries into semi-structured layouts and then iteratively refines those representations using defender feedback. The method is framed as an extensive-form game with incomplete information and a Hidden Chain-of-Table-Thought mechanism, with the interaction described as moving toward pooling and Subgame Perfect Nash equilibria. The empirical evaluation uses FSBench, containing 494 queries across six financial-risk categories, and compares SO-STAG with several single-turn and multi-turn jailbreak baselines across six target-model columns. Additional analyses study iteration budget, defensive mechanisms, auxiliary-generator scale, evaluator consistency, and computational efficiency. The paper reports a 97.78% average attack success rate and substantially fewer input tokens than the two semantic multi-turn baselines. Its main contribution is the empirical observation that structured representations can form a strong attack surface, together with an automated feedback-driven procedure for exploring that vulnerability.

Critical strengths
  • S1: The paper focuses on a useful and underexplored safety question: whether harmful financial intent can evade alignment through information structure rather than primarily through lexical or narrative transformations. The distinction between semantic concealment and structural concealment is practically relevant for systems that increasingly consume tables, ledgers, and semi-structured financial records, and it gives the paper a clear research target.
  • S2: The empirical attack performance is strong across the model columns reported in Table 2. SO-STAG reaches very high ASR values on all six evaluated targets, and the comparison includes both single-turn attacks and strong multi-turn methods such as Crescendo and RCA. Even setting aside the paper's broader theoretical interpretation, these results suggest that the proposed attack procedure is worth taking seriously as a red-teaming tool.
  • S3: The ablation in Table 6 is informative because it separates the contribution of one-shot structural obfuscation, multi-turn refinement without feedback, and the full feedback-driven procedure. The increase from 44.78% to 88.63% and then to 99.81% on Qwen3 provides useful evidence that iterative interaction and refusal-aware feedback are important to the final attack success, rather than the result being attributable only to the initial formatting transformation.
  • S4: The paper goes beyond a single headline ASR number. Table 4 studies the interaction budget, Table 5 tests several defense variants, Table 7 varies the auxiliary generator, and Table 8 compares multiple automated judges with human evaluation. These analyses address several plausible alternative explanations for the attack's performance and are useful directions for evaluating an adversarial framework.
Critical weaknesses
  • W1: My main concern is the internal consistency of the core evaluation. Several values in the Avg column of Table 2 do not appear to be the arithmetic mean of the six displayed model-specific ASRs. In addition, Appendix A.7 lists Qwen2.5-72B and GPT-OSS-120B among the evaluation targets, whereas Table 2 contains Qwen2.5-32B and LLaMA-4-Scout. Appendix A.3 states that a unified LLaMA-3.1 judge with temperature 0 is used, while Appendix A.7 describes GPT-OSS-120B as the judge with temperature 0.01. These discrepancies make it difficult to determine exactly which experimental pipeline produced the headline results.
  • W2: The central causal claim about structural blindness is not cleanly isolated from semantic and agentic optimization. The Introduction states that the same malicious intent is hidden at essentially zero semantic cost, but Algorithm 1 uses generative policies and iterative defender feedback to construct and revise the attack. Table 6 ablates the number of interaction rounds and the use of feedback, but it does not compare matched prompts in which semantic content is held fixed while only the structural representation changes. As a result, the current evidence convincingly shows that the full SO-STAG pipeline is an effective jailbreak procedure, but it is less clear that the gains can specifically be attributed to structural topology rather than to a combination of structural rewriting, semantic reframing, and adaptive search.
  • W3: I am not convinced by the game-theoretic claims in Secs. 3.1-3.3. The paper repeatedly states that structural masking induces a pooling equilibrium and that the iterative process converges to a Subgame Perfect Nash Equilibrium, but I could not find a defender utility, an explicit belief-update rule, an equilibrium derivation, or a proof of existence/convergence. Algorithm 1 treats the empirical stopping condition SUCCESS = True and R(y) <= epsilon as indicating that an SPNE has been achieved. That condition is not, by itself, sufficient to establish sequential rationality in every subgame. Similarly, the defender's inability to observe a hidden reasoning trace does not by itself establish a pooling equilibrium. The empirical method does not require these strong theoretical claims, so either a more complete formal analysis or a substantially narrower interpretation would improve the paper.
  • W4: The fairness of the baseline and efficiency comparisons is not sufficiently clear. Table 2 says that multi-turn attacks use a maximum budget of five dialogue rounds, whereas Appendix A.3 describes RCA with up to eight iterations and ReNeLLM with up to ten iterations; Crescendo is set to five turns. Table 3 then compares token counts, runtime, and peak ASR across methods that appear to use different attack generators, iteration limits, and success levels. The claim of more than 80% fewer input tokens is supported relative to Crescendo and RCA, but the paper does not yet make clear whether the token/runtime comparison is measured under a common success target and common interaction budget. Since efficiency is one of the paper's headline claims, the accounting protocol should be unambiguous.
  • W5: The paper provides little uncertainty information for comparisons whose margins can be small. In Table 2, the average difference between SO-STAG and RCA is about one percentage point, and on LLaMA-4-Maverick the displayed numbers differ by only 0.02 points. I could not find repeated runs, confidence intervals, or another estimate of variability. Table 8 is a useful evaluator check, but the manuscript does not specify the number of human annotators, inter-annotator agreement, adjudication procedure, or whether annotation was blinded. These details matter because the paper makes strong statements that the evaluation is free from proxy inflation and that small differences reflect real performance gains.
  • W6: The construction of FSBench is not documented consistently enough for a new benchmark that supports the paper's domain-generalization claims. Appendix A.7 says that the dataset integrates three complementary data streams but explicitly describes only the first stream before moving to the agent configuration. Figure 3 gives source proportions, while the surrounding text does not clearly reconcile those proportions with the stated 223 samples from prior benchmarks. I also could not find a clear description of deduplication, overlap handling, or how generated samples were validated. These issues do not invalidate the attack results, but they make the scope and independence of the benchmark harder to assess.
Questions for authors
  • Q1. How is the Avg column in Table 2 computed? If it is intended to be the mean over the six displayed target-model ASRs, could you reconcile the values for ArtPrompt, ReNeLLM, CodeChameleon, and Crescendo with the corresponding row entries and provide corrected numbers if necessary?
  • Q2. Could you provide a single definitive specification of the target models and judging pipeline used for each main result? In particular, how should I reconcile Qwen2.5-32B in Table 2 with Qwen2.5-72B in Appendix A.7, the presence of LLaMA-4-Scout in Table 2 versus GPT-OSS-120B in the Appendix target list, and the LLaMA-3.1 judge at temperature 0 in Appendix A.3 versus the GPT-OSS-120B judge at temperature 0.01 in Appendix A.7?
  • Q3. What evidence isolates the effect of structural topology from semantic rewriting? For example, are the original harmful query and the final structured query constrained to preserve the same wording or semantic content, and is there a matched control comparing flat versus hierarchical/tabular representations under otherwise identical content? A convincing answer here would substantially strengthen the paper's central claim of structural rather than general adaptive evasion.
  • Q4. What formal argument establishes the claimed pooling equilibrium and SPNE in Secs. 3.1-3.3? Specifically, what is the defender's utility, how are beliefs over attacker type updated, and why does the stopping condition in Algorithm 1 imply an SPNE rather than simply a successful empirical attack? If the equilibrium language is intended mainly as an interpretation, would the authors be willing to narrow those claims?
  • Q5. Are the Table 2 and Tables 4-7 numbers based on repeated independent runs? If so, could you report variability or confidence intervals, especially for comparisons with RCA where the margins are small? For Table 8, how many human annotators were used, what rubric and adjudication procedure were followed, and was inter-annotator agreement measured?
  • Q6. Could you clarify the complete construction of FSBench, including all three data streams mentioned in Appendix A.7, their exact counts, deduplication/overlap handling, and the validation procedure for generated queries? This would help determine whether the reported cross-category robustness reflects genuinely diverse test cases rather than repeated or closely related prompts.
Reviewer_1yFQOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper focuses on structural evasion in financial safety alignment and proposes SO-STAG, an attack framework based on structural re-encoding, multi-turn interaction, and judge-feedback optimization. The main value of the paper lies in pointing out that existing safety guardrails may suffer not only from semantic vulnerabilities, but also from a lack of robustness to risk expressions embedded in tables, hierarchical layouts, and semi-structured representations. The empirical results are overall strong: the method achieves high attack success rates across multiple target models and multiple categories of financial risk, while also showing better efficiency than several multi-turn baselines. However, I am not yet convinced by the current version because the problem setting is quite narrow, the central claim about structural vulnerabilities is not sufficiently supported by clean controlled experiments, and the generalization analysis remains limited. For these reasons, I find the paper interesting but not yet strong enough in its current form.

Critical strengths
  • The empirical results are overall strong. The authors report high ASR across multiple target LLMs and multiple categories of financial risk, while also showing efficiency advantages over the main multi-turn baselines. In particular, the multi-model and multi-category experimental setup makes the results look less like an isolated phenomenon.
  • The contribution of the multi-turn feedback mechanism is supported by empirical evidence. The ablation studies distinguish between single-turn, multi-turn without feedback, and the full setting, which more directly demonstrates the importance of multi-turn interaction and judge feedback for the final performance gain. This is more convincing than simply reporting the main results.
Critical weaknesses
  • The problem setting is too narrow. In my view, studying attacks on structured data is already fairly narrow, and further restricting the setting to finance makes it even narrower. As foundation models continue to improve, I suspect that models may naturally become better at handling such cases, which reduces the significance of studying a problem with such limited scope. For example, the models used in the experiments are all more than a year old, while the performance of open-source LLMs has improved substantially since then. It is possible that this problem may no longer persist to the same extent.
  • The central “structural” claim lacks cleaner controlled experiments. Although the paper emphasizes that the method exploits a structural blind spot, it does not systematically compare different structural patterns, nor does it directly compare “semantic rewriting” and “structural rewriting” under the same multi-turn framework. As a result, it remains unclear whether the performance gains truly come from structural representations themselves, or mainly from the multi-turn agent + feedback design.
  • The generalization evaluation is still insufficient. The paper evaluates cross-model, cross-category, and cross-judge performance, but it does not answer a more fundamental question: would the method remain effective if the attack pattern or structural template were changed? The current experiments mainly support the claim that the method works under the specific tabular/hierarchical patterns used in the paper, but they are not sufficient to support a broader conclusion about “structural blind spots” in general.
Questions for authors

See above.

Reviewer_kS19OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposed Structural Risk Obfuscation Attack (SO-STAG), a red-teaming framework that rewrites harmful financial requests into tabular/hierarchical "structured" prompts and refines them using a judege LLM. Experiments on FSBench shows 97.78% average attack success rate over 6 LLMs.

Critical strengths
  1. Jailbreaking in the financial domain is an interesting problem, and this paper organizes it into six misconduct categories.
  2. The paper compares with seven single-turn and multi-turn baselines, and also includes judge–human agreement.
Critical weaknesses
  1. The main paper has no experiments. Section 4 only has one paragraph that lists the baselines, and all the experiments are in the appendix. For the claims in the abstract and conclusion, such as the 97.78% average evasion success rate, I cannot find any support in the main paper. It really hurts the readability.

  2. The figures are hard to read. For example, Figure 1's caption refers to left and right, but the figure itself has three rows. The "structured tabular representations and hierarchical gray-area prompts" mentioned in the caption do not even appear in the figure, and the image in the third row is too small to read. Figure 2 is blurry, looks like a collage of screenshots, and mixes different fonts and font sizes.

  3. Some sentences look AI-generated, e.g, "Your proposed Hidden Chain-of-Table-Thought (Hidden CoTT)..." in lines 192-193.

  4. The results are internally inconsistent, even in the appendix. For example, the "Avg" column in Table 2 does not match the row values for four of the eight methods. The set of "six leading LLMs" also changes across the paper: Table 2 uses Qwen2.5-32B, but Table 4 uses Qwen2.5-72B. The text in appendix says every model reaches ≥95% by T=3, but Table 4 shows 94.81 on GLM-4.7 and 94.13 on LLaMA-4-Mav.

  5. The paper has a lot of typos, such as double ")", using wrong quotation marks, and missing spaces between the text and citations.

  6. The game-theoretic part is only notation, e.g., the claimed pooling equilibrium and SPNE are never derived, the defender's payoff is not defined, λ\lambda and ϵ\epsilon are not reported, and ReEncode, LatentReason and TableRep are not specified.

  7. The paper claims violations are camouflaged "without falsifying the underlying factual content", but I couldn't find anywhere in the paper where this claim is measured.

Questions for authors
  1. How can results such as 99.81% (from table 2) and 99.87% (from table 5) arise from 494 queries? 493/494 is already 99.80%.
  2. Which six models were evaluated? Table 2 uses Qwen2.5-32B but Table 4 uses Qwen2.5-72B.
  3. How is the "Avg" column in Table 2 computed? It does not match the row values for ArtPrompt, ReNeLLM, CodeChameleon and Crescendo.
  4. How do the authors check that the obfuscated prompts keep the facts of the original queries?
  5. What values of λ\lambda and ϵ\epsilon are used? Can the authors show example prompts generated by SO-STAG?
  6. Is any equilibrium actually derived or checked, or is the game-theoretic part only a description of the algorithm?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.