- 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.