acceptodds
Under review as a conference paper at ICLR 2027

CDIM: Cross-Domain Isomorphic Mapping for Efficient Black-Box Jailbreak Attacks

Abstract

Jailbreak attacks are an effective paradigm for testing the safety boundaries of Large Language Models (LLMs), exposing vulnerabilities to facilitate the development of robust defenses. Existing black-box jailbreak methods primarily rely on search-based prompt optimization or long-context construction, which typically suffer from high query overhead and insufficient stealth when targeting frontier models. To address these limitations, we propose Cross-Domain Isomorphic Mapping (CDIM), a novel black-box jailbreak framework that conceals and executes malicious intents through structurally analogous yet semantically benign carriers. By recoding sensitive requests into Out-of-Distribution (OOD) expressions that are superficially harmless but logically isomorphic, CDIM systematically improves attack stealthiness while minimizing query costs. Mechanistically, CDIM operates in three stages: (1) Mapping Construction, which selects a safe semantic domain and builds a bijective dictionary between sensitive semantics and benign carriers; (2) Isomorphic Simulation, which restricts the target model's generation within the safe semantic space while preserving the original procedural logic; and (3) Backward Recovery, which decodes the implicit benign output back into actionable sensitive content. Extensive experiments on two mainstream jailbreak benchmarks demonstrate CDIM's superior efficiency and effectiveness. Notably, with merely three queries, CDIM achieves an average attack success rate of 65.2% against Claude Sonnet 4.5 , significantly outperforming existing baselines.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Jailbreak Attack, Cross-Domain Isomorphic Mapping, Black-box Attack

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_53AvOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes CDIM, a three-stage black-box attack framework that rewrites harmful requests into alternative semantic domains, queries a target model, and uses an auxiliary model to generate the final answer from the target response and original request. Experiments on AdvBench and miniHarmBench report improved end-to-end attack success and query efficiency.

Critical strengths

None

Critical weaknesses
  1. The experiment evaluation fails to establish that the target models were actual jailbroken. ASR is computed on answers generated by a separate recovery model that receives the original harmful query and can provide the harmful information itself. The results show that the full pipeline can produce harmful answers, but fail to show that the critical harmful information is actually provided by the target model.

  2. The “isomorphic” property is neither enforced nor independently validated. Matching entity names does not ensure that the rewritten task preserves the original rules and constraints. A correct final answer cannot establish this either, because the recovery model may fill in information lost during rewriting.

Questions for authors
  1. How does performance change when the target response is removed or replaced with unrelated text, holding other inputs and the recovery budget fixed? Please provide complete representative traces identifying which critical harmful information originates in the target response.

  2. How is the relation-preservation condition in Equation 3 independently verified? What evidence shows that preserving this structure improves performance over ordinary semantic rewriting under matched settings?

  3. Why does CDIM use Qwen3-Max while the baselines use Mixtral with reduced search settings? Do the reported gains persist with the same auxiliary models and comparable search budgets?

Reviewer_Uv8bOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Existing black-box jailbreak methods often rely on iterative prompt search, multi-turn interaction, or long-context construction, which can incur high query costs, expose malicious intent, or distort task semantics during transformation. To address these limitations, this paper proposes CDIM, a cross-domain isomorphic mapping framework. CDIM maps the entities, relations, and task constraints in a malicious request into superficially benign carrier domains, such as code, magic, or mathematics. The target model then produces an intermediate response in the carrier domain, and a recovery model generates the final answer using that response, the mapping dictionary, and the original request. The authors evaluate CDIM on AdvBench and miniHarmBench across multiple open- and closed-source models. The results show that the full pipeline achieves high end-to-end attack success rates against several strongly aligned models and outperforms multiple baselines under the reported experimental settings.

Critical strengths

S1. The attack idea is intuitive and practically relevant. By mapping malicious requests into superficially benign carrier domains such as code, magic, or mathematics and subsequently recovering the intermediate response, the paper constructs a clear cross-domain attack pipeline. The approach highlights an important risk in multi-model workflows: inspecting only the input to a single stage may be insufficient to identify malicious intent distributed across the pipeline.

S2. The experimental coverage is relatively broad. The paper evaluates multiple open- and closed-source target models on AdvBench and miniHarmBench and compares three carrier domains: Code, Magic, and Math. This breadth provides a useful view of how the method behaves across model families and carrier designs.

S3. The discussion of limitations is relatively transparent. The main text and appendix explicitly acknowledge challenges including information-source attribution, mismatched baseline and auxiliary-model configurations, the limited scope of query-count statistics, and the absence of rigorous verification of structural isomorphism. This transparency helps readers interpret the scope of the reported results.

Critical weaknesses

W1. The experiments do not establish that the target model itself is jailbroken. The final responses used to compute attack success are generated by the recovery model, which directly receives the original malicious request. Consequently, a high attack success rate demonstrates that the complete multi-model pipeline can produce harmful outputs, but it does not establish that the target model itself has been jailbroken or that the harmful information in the final response originates from the target model's intermediate output. This creates a substantial gap between the reported evidence and the paper's central claim about jailbreaking the target model.

W2. Target outputs lack carrier-domain constraints and recoverability guarantees. The target model generates text freely, without constraints requiring it to use entities from the mapping dictionary, remain within the predefined carrier domain, preserve all source relations and constraints, or support unique inverse recovery. The method also provides no explicit detection or fallback mechanism when the output contains unmapped content, omits key constraints, or deviates from the carrier semantics. The proposed cross-domain mapping therefore lacks a reliable closed-loop guarantee.

W3. The ablation study is insufficient to identify the source of information. The component ablations show that the full system performs better than reduced variants, but they do not reveal which information from the target output is actually used by the recovery model. In particular, the recovery-only condition still gives the recovery model access to the original malicious request and therefore does not constitute a clean control. The current experiments cannot distinguish whether success is driven by useful information from the target model, independent generation by the recovery model, or their joint effect.

W4. Inconsistent baseline configurations weaken the fairness of the comparison. CDIM uses the relatively strong Qwen3-Max model for mapping and recovery, whereas the baselines use Mixtral and reduced search configurations. The reported performance gaps are therefore confounded by differences in auxiliary-model capability and attack budget and cannot be attributed solely to CDIM's method design. Without matched auxiliary models, call budgets, and retry policies, the claim of superiority over existing methods is not sufficiently supported.

W5. The query-efficiency comparison uses an inconsistent accounting scope. The headline claim of three queries corresponds to a no-retry execution with one call to each of the mapping, target, and recovery models. It does not characterize retries, failed runs, token consumption, latency, or monetary cost. Moreover, some baselines primarily count target-model queries, whereas CDIM shifts substantial computation to auxiliary models. Without a unified cost ledger, the claimed efficiency advantage cannot be reliably established.

Questions for authors

Q1. Can you provide an ablation study where the recovery model generates the final response without access to the original malicious request, to prove the harmful information genuinely originates from the target model's intermediate output?

Q2. What specific automatic detection or fallback mechanism does your system employ to guarantee reliable inverse recovery when the target model's output deviates from carrier semantics or contains unmapped entities?

Q3. Can you rerun the comparison between CDIM and the baselines using strictly matched auxiliary models and identical call/retry budgets to isolate the methodological contribution from model capability differences?

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

The paper proposes CDIM, a black-box jailbreak pipeline that uses an auxiliary model to rewrite a sensitive request into a benign-looking carrier domain and construct an entity dictionary. The target answers the transformed request, after which an auxiliary recovery model generates an original-domain answer using the target response, inverse dictionary, and original request. Code, Magic, and Math carriers are evaluated across eight target configurations on AdvBench and miniHarmBench. The paper reports high end-to-end success rates and improvements over component removals. Its relational-preservation condition is a specification, not an established property of the generated transformations.

Critical strengths

Coverage and component controls. The experiments cover two benchmarks, eight target configurations, and three carrier domains. Table 5 includes a recovery-only comparison, which usefully exposes the substantial success already achievable with that component.

Critical weaknesses
  • Recovery is not verified as translation without substantive additions. To establish that the target supplies recoverable jailbreak content, recovery should only translate information already represented in the target response under the declared mapping. However, recovery performs conditional generation with access to the original malicious request (Section 3.4, Equation 5), and the success metric scores its final answer (Section 3.1). Table 2 explicitly acknowledges differences beyond inverse dictionary lookup. The evaluation does not rule out the recovery model supplying substantive information absent from the target response. Full-pipeline improvements over recovery-only results in Table 5 do not establish the provenance of that information, a limitation acknowledged in Appendix D. Without verified content-preserving recovery, the reported success does not convincingly establish that the target supplied the jailbreak content.
  • The mapping does not establish preservation of the original task's semantics. In Table 1, the original task is to establish and operate a marketplace selling illegal drugs. The transformed prompt instead asks for a fantasy marketplace trading "Ethereal Blossoms": selling illegal drugs is no longer the task described to the target. The paper does not establish that the original task remains semantically encoded in this prompt, rather than being replaced by a different task. It therefore remains unclear how answering the fantasy request elicits useful information about selling illegal drugs, even in encoded form. If the target response lacks that information under the mapping, faithful translation cannot supply it; recovery would have to add substantive content to fulfill the original request. Entity correspondences and shared commercial structure do not resolve this gap. Equation 3 specifies a desired preservation condition, but is explicitly not a theorem about the implemented prompts. The paper does not convincingly establish that the target generates meaningful content fulfilling the original malicious task, in any format.
  • The comparisons do not isolate an advantage of the proposed method. CDIM uses Qwen3-Max as its default auxiliary model, whereas baselines use Mixtral with reduced configurations (Section 4.1 and Appendices A.2 and G). Appendix G explicitly notes that auxiliary capability and budgets are not matched. The reported differences therefore compare differently configured systems and do not isolate the benefit of isomorphic transformation over existing obfuscation or contextual approaches. This is particularly important when an auxiliary model, rather than the target, produces the scored answer.
Questions for authors
  • Recovery fidelity: Can the reported success be reproduced with recovery verified to perform only content-preserving translation under the declared mapping? Please provide complete paired target and recovered responses showing that substantive information in the final answer was already represented in the target response, rather than added during recovery.
  • Task preservation: In Table 1, what conveys the original task of selling illegal drugs through the fantasy-marketplace prompt, rather than replacing it with a different task? Please demonstrate that the target's response contains useful information fulfilling the original request under the mapping, and provide evidence of such semantic preservation across the evaluated cases.
  • Controlled comparisons: Do the reported advantages persist with matched auxiliary-model capability and justified baseline configurations? Please distinguish the transformation's contribution from auxiliary-model and resource differences, reporting target and auxiliary costs separately.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.