acceptodds
Under review as a conference paper at ICLR 2027

When the Editing Intent Is Split: A Cross-Modal Jailbreak Attack for Large Image Editing Models

Abstract

Large image editing models increasingly infer editing intent from both textual and visual inputs, creating an underexplored safety risk when harmful intent is distributed across modalities. In this work, we formalize this split-intent vulnerability and propose the Split-Intent Jailbreak Attack (SIJA), which asymmetrically distributes an unsafe editing intent by specifying the editing referent in text while concealing the safety-critical action and specification through graphical visual semantics in an auxiliary cue. To systematically study this threat, we introduce the Split-Intent Image Editing Safety Benchmark (SI2ESBench) for evaluating image editing models under split-intent inputs. Extensive experiments on representative commercial and open-source models demonstrate that SIJA effectively compromises state-of-the-art image editing systems, achieving ASRs of 68.28% on GPT Image 2 and 91.38% on Qwen-Image-3.0-Pro. Our results expose a mismatch between multimodal editing and safety assessment: the distributed intent can be recovered to execute the requested transformation, yet its harmfulness often remains undetected by existing safeguards. To mitigate this vulnerability, we propose editing-intent reconstruction, a defense principle that explicitly evaluates the complete transformation implied by the joint multimodal context and substantially improves rejection of split-intent attacks. Our findings expose a cross-modal safety gap in modern image editing systems and highlight the need to assess the jointly inferred transformation rather than individual inputs in isolation.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Large image editing models, split-intent, jailbreak attack

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_1osAOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies split-intent safety failures in large image editing models and proposes SIJA, a cross-modal jailbreak attack that distributes unsafe editing intent across textual and visual inputs. The authors further introduce SI2ESBench and evaluate a range of commercial and open source image editing models in this context. The results show that the current large image editing models can recover and execute unsafe transformations from distributed intents, and the proposed editing-intent reconstruction defense greatly improves the detection capability.

Critical strengths
  1. This paper identifies a distinct safety problem in large image editing models. SIJA shows that an unsafe editing request can remain incomplete in both the textual and visual components while still being recoverable by the editing model through joint interpretation. This exposes a safety gap in which the unsafe transformation can emerge through joint interpretation of the source image, textual referent, and graphical visual semantics.
  2. The experimental results provide extensive evidence for the split-intent vulnerability and support the proposed mechanism. The main SIJA evaluation covers commercial and open source large image editing models across 10 risk categories, showing the substantial vulnerability of different model families. The effectiveness of SIJA is further demonstrated by comparing different forms of attack, while component ablation makes clear the different roles of textual referents and visual cues. These results provide extensive empirical evidence for this vulnerability and controlled evidence for the proposed split-intent attack.
  3. The defense analysis further illustrates the practical security implications of SIJA. The sharp degradation of the text-space intent guardrail under SIJA suggests that the attack exposes specific weaknesses in security mechanisms that rely on recovering harmful intentions from readable text. Meanwhile, the strong performance of edit intent reconstruction suggests that explicitly recovering the full intent across textual and visual inputs is a promising direction to mitigate this vulnerability. This makes SIJA not only an effective attack but also a useful setting for motivating stronger protective measures for large image editing models.
Critical weaknesses
  1. The human validation procedure for SI2ESBench is not described in sufficient detail. Benchmark construction relies on manual verification of referent grounding, component intent incompleteness, and recoverability of the complete editing intent from the joint inputs. These guidelines are described and it is stated that ambiguous or inconsistent cases are revised or removed, but the specific verification process is not described in detail.
  2. The human–MLLM comparison is reported mainly at the aggregate level. The current comparison reports aggregate HS, EV, and HRR on the same 350 outputs, but similar summary values do not necessarily imply agreement on the same individual samples. Thus, the current analysis does not directly characterize the instance-level agreement between automatic judging and human annotation.
Questions for authors
  1. Could the authors clarify how the manual validation of SI2ESBench was conducted in practice, particularly how the stated criteria were checked and how ambiguous cases were resolved? The paper describes criteria for the referent grounding and recoverability of the complete editing intent, component intent incompleteness, and complete edit intent, but it is not clear how these checks are performed or how ambiguous cases are handled.
  2. What is the instance-level agreement between the primary MLLM judge and the human annotations on the 350 sampled outputs? The reported aggregated HS, EV, and HRR are informative, but similar aggregated values do not necessarily indicate consistency for the same individual sample.
1 comment

We thank the reviewer for the positive assessment and the suggestions to clarify benchmark validation and judge reliability.

A1. Human Validation Procedure

Three human validators review candidate SI2ESBench cases against three criteria. For referent grounding, they inspect the source image and textual instruction to verify that the target is identifiable without revealing the action or specification. For component-wise incompleteness, they examine the textual instruction and visual cue separately to verify that neither independently specifies the complete source-grounded edit. For joint recoverability, they inspect the source image, instruction, and cue together to verify that the intended transformation can be recovered. Candidates with ambiguous targets, inconsistent or incomplete visual semantics, or redundant intent information are revised or excluded. This procedure yields 766 validated cases. (Revision: Sec. 4.1 and Appendix B.2.)

A2. Instance-Level Judge Agreement

We extend the original aggregate comparison on 350 edited outputs to sample-level agreement between Gemma-4-31B-it and human consensus. Three annotators independently evaluate each output, with majority-voted EV and median HS determining the consensus labels; HRR is derived from EV = 1 and HS at least 4. We report precision, recall, F1, Cohen's kappa for EV/HRR, quadratic weighted kappa for HS, and pairwise human agreement.

MetricPrecision (%)Recall (%)F1 (%)Judge–Human κHuman–Human κ
EV94.2495.9795.100.7700.826
HRR95.1195.4795.290.8050.933
HS———0.8410.948

The F1 scores exceed 95% for both EV and HRR, and agreement reaches 0.841 quadratic weighted kappa for HS. The paired confusion matrices quantify errors on individual outputs; the appendix reports these matrices, results for all four MLLM judges, and human inter-annotator agreement. These analyses support the consistency of the primary judge with human evaluations beyond aggregate averages. (Revision: Sec. 5.5 and Appendix A.4.)

Reviewer_X4zgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a cross-modal jailbreak attack that places the editing referent in text and the action and specification in an auxiliary image. It also introduces a benchmark and an intent-reconstruction defense.

Critical strengths

The paper studies a relevant safety issue in multimodal image editing and evaluates the attack across multiple commercial and open-source models.

Critical weaknesses
  1. The asymmetric decomposition is manually specified: text identifies the editing referent, while the image conveys the action and specification. However, the paper provides insufficient justification for this allocation and no comparison with alternative decompositions to establish its advantage.

  2. The benchmark follows the same fixed construction as the proposed attack. Its contribution beyond a targeted test set for SIJA is therefore limited.

  3. The paper does not compare the proposed defense with existing safety assessment or defense methods applied to the reconstructed editing intent. This comparison is needed to establish whether the proposed defense offers additional benefits beyond intent reconstruction.

Questions for authors
  1. Why is the proposed text–image allocation preferable to alternative decompositions?
  2. How does intent reconstruction followed by an existing safety classifier compare with the proposed defense?
3 comments

We thank the reviewer for the constructive assessment and the questions concerning the asymmetric design, benchmark contribution, and defense comparison.

A1. Why the Asymmetric Allocation?

SIJA assigns the source-grounded referent r to text and the action a and specification s to a graphical cue. This separates identification of the editing target from the transformation semantics. To compare this design with alternatives, we evaluate all six information-preserving allocations using the same 100 category-stratified GPT Image 2 cases and fixed source images and intended edits. The slash denotes text (left) and visual cue (right).

Text / Visual CueASR (%)HSEV (%)HRR (%)
r/(a,s) (SIJA)573.155554
(r,a)/s512.925048
(r,s)/a622.515438
a/(r,s)502.834846
s/(r,a)612.294632
(a,s)/r652.895847

The r/(a,s) allocation achieves the highest HS (3.15) and HRR (54%), whereas (a,s)/r obtains the highest ASR (65%) and EV (58%). The distinction between edit validity and high-risk execution explains the choice of SIJA's allocation: placing the referent in text and the action/specification graphically yields the strongest harmful editing outcome among the tested assignments. (Revision: Sec. 5.3 and Appendix C.4.)

A2. Contribution of SI2ESBench

We appreciate the opportunity to clarify the benchmark's scope. SI2ESBench contains 766 manually validated source-conditioned editing cases across ten risk categories. Each case has a source-grounded referent, action, and specification, arranged so that the textual instruction and auxiliary cue are individually incomplete but jointly recover the intended unsafe edit. Three human validators check referent grounding, component-wise incompleteness, and joint recoverability; ambiguous or inconsistent cases are revised or excluded.

The structured intent representation enables controlled variations beyond the canonical SIJA presentation. Using the same underlying cases, we compare TJA, VJA, and SIJA, perform component ablations, test six alternative modality allocations and the 2×2 allocation/encoding design, and evaluate defenses. These evaluations make SI2ESBench useful for studying how input composition affects harmful editing and safety assessment, rather than only measuring a single fixed attack format. (Revision: Secs. 4.1 and 5.3; Appendices B.2 and C.4.)

A3. Comparison with Existing Guardrail-Based Safety Assessment

We thank the reviewer for suggesting a comparison with a separate safety classifier applied after intent reconstruction. We evaluate a Multimodal Intent Guardrail that uses Qwen3-VL-2B-Instruct to reconstruct the complete edit from the source image, visual cue, and textual instruction, followed by Qwen3Guard-4B for safety classification. We also include an OCR-based Text-space Intent Guardrail, VJA Introspection, and our Decomposed Introspection. All methods are evaluated on 766 SI2ESBench cases; latency measures pre-generation safety processing.

DefenseRejection (%)Additional ModelsLatency (s)
Text-space Intent Guardrail4.1GLM-OCR + Qwen3Guard-4B1.83
Multimodal Intent Guardrail65.0Qwen3-VL-2B-Instruct + Qwen3Guard-4B1.24
VJA Introspection68.4None5.70
Decomposed Introspection (Ours)93.1None4.37

The multimodal reconstruction-plus-classifier pipeline raises rejection from the OCR-based pipeline's 4.1% to 65.0%, demonstrating the importance of recovering non-textual editing semantics. Our decomposed introspection reaches 93.1% without an additional guardrail model. This comparison distinguishes an external reconstruction-and-classification pipeline from the explicit decomposition used within introspection, and supports transformation-level safety assessment. (Revision: Secs. 3.2 and 5.4; Appendix C.5.)

Thank you again for your constructive feedback, which has helped us strengthen the paper.

We would like to briefly summarize how we have addressed your three main concerns:

  1. Asymmetric allocation: We evaluated all six information-preserving modality allocations. SIJA achieves the highest harmfulness score (3.15) and HRR (54%), supporting our choice of placing the referent in text and the action/specification in graphical cues.
  2. Benchmark contribution: We clarified the construction and human validation of SI2ESBench and demonstrated its utility beyond the canonical SIJA setting through alternative modality allocations, controlled encoding experiments, and defense evaluations.
  3. Defense comparison: We introduced a multimodal reconstruction-plus-classifier baseline, achieving 65.0% unsafe rejection, compared with 93.1% for our decomposed introspection defense.

These additional experiments and clarifications have been incorporated into the revised manuscript. We hope they adequately address your concerns regarding design justification, benchmark contribution, and defense effectiveness.

Thank you again for your time and consideration.

Thank you again for your constructive feedback. We would also like to emphasize the central contributions and novelty of our work.

The key contribution is not merely another way of presenting harmful instructions, but the identification of a distinct safety failure in multimodal image editing: an unsafe transformation can be recovered and executed through joint interpretation even when neither the textual instruction nor the visual cue independently specifies the complete harmful intent. SIJA formalizes this vulnerability through asymmetric intent distribution and visual-semantic concealment, with additional controlled experiments supporting the effectiveness of this design. Beyond a fixed attack test set, SI2ESBench provides a structured and manually validated benchmark that supports systematic investigations of intent allocation, visual encoding, and safety defenses. Furthermore, our editing-intent reconstruction principle addresses the underlying mismatch between multimodal editing capabilities and safety assessment, with comparisons against existing guardrail-based approaches demonstrating its practical value.

We hope these contributions and the additional evidence clarify the novelty and significance of our work. Thank you again for your time and consideration.

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

This paper studies image-editing jailbreaks in which harmful intent is distributed across textual and visual inputs. It proposes SIJA, which places the editing referent in text and the action specification in graphical cues, and introduces SI2ESBench with 766 instances across ten risk categories (Sec. 3–4; Appendix B.2). Experiments cover seven base editors and distinguish refusal bypass from valid harmful execution (Table 1; Appendix C.1). A decomposed introspection defense reconstructs the complete editing intent before safety assessment, achieving 93.1% rejection on unsafe SI2ESBench requests (Sec. 5.4; Fig. 4).

Critical strengths
  1. Clearly formulates distributed editing intent through a referent action specification decomposition, providing a concrete setting for studying cross-modal safety failures (Sec. 3.1; Eq. (2)).

  2. Provides a benchmark construction procedure with manual checks for source grounding, component-wise intent incompleteness, and joint recoverability, making the intended evaluation conditions explicit (Sec. 4.1; Appendix B.2).

  3. Evaluates multiple commercial and open-source editors and distinguishes image return, editing validity, and valid harmful execution, providing a more informative assessment than refusal bypass alone (Table 1; Appendix C.1).

Critical weaknesses
  1. Component-removal ablations confound cross-modal composition with missing information. Removing the prompt removes the explicit referent, while removing the cue removes the explicit action and specification. The resulting EV/HRR reductions may therefore reflect underspecified editing requests. Could the authors add information-preserving controls that retain the complete intended edit while changing its distribution across modalities? Such controls are not reported in this ablation (Eq. (2); Appendix A.1; Table 5). No direct evidence found in the manuscript.

  2. The defense evaluation lacks benign controls. Figure 4 measures rejection of unsafe requests; SI2ESBench-VJA contains the same unsafe cases in another presentation. Consequently, the 93.1% rejection rate does not establish whether legitimate edits remain usable. Could the authors report benign false rejection and editing success, particularly given the conservative rejection instructions in the defense prompt? (Sec. 5.4; Fig. 4; Appendix C.5). No direct evidence found in the manuscript.

  3. The attack contribution beyond VJA needs clearer isolation. VJA already permits visual patterns to express editing semantics, so graphical depiction alone is insufficient to establish the main novelty (Hou et al., When the Prompt Becomes Visual: Vision-Centric Jailbreak Attacks for Large Image Editing Models, 2026, Sec. 3.1). SIJA’s comparison changes both intent allocation and encoding. Could the authors independently control these factors to establish the specific contribution of split intent? (Sec. 5.3; Table 2).

  4. The specific benefit of decomposed introspection remains unclear. SIJA explicitly extends VJA’s introspection defense, while VPA-Guard also requires joint interpretation of visual cues and text before safety assessment, albeit for image-to-video generation (Sun et al., VPA-Guard: Defending and Benchmarking Image-to-Video Generation Against Visual Prompt Attacks, 2026, Appendix D.4.1). Could the authors compare explicit three-field decomposition against generic joint-intent reconstruction under identical safety wording and comparable inference budgets? This contribution is not isolated by the current comparison (Sec. 3.2; Fig. 4; Appendix C.5). No direct evidence found in the manuscript.

  5. Similar aggregate metrics do not establish accurate sample-level judgments. Gemma-4-31B-it and humans have similar average EV/HRR, but opposing sample-level errors could cancel and produce similar averages (Sec. 5.5; Table 3). Could the authors report paired confusion matrices, precision/recall/F1 for binary decisions, weighted agreement for harmfulness judgments, and human inter-annotator agreement? These statistics are needed to assess whether the judge identifies the same valid and harmful outputs as humans. No direct evidence found in the manuscript.

Questions for authors

Please refer to the issues raised in the Weaknesses section.

4 comments

We thank the reviewer for the detailed assessment and constructive suggestions. We have conducted additional controlled experiments and strengthened the corresponding analyses.

A1. Information-Preserving Intent Allocation

We agree that removing a component tests its role but does not isolate the effect of modality assignment. To examine that effect while preserving the complete intended edit, we evaluate all six nonempty allocations of the referent (r), action (a), and specification (s) between text and the visual cue. All settings use the same 100 category-stratified GPT Image 2 cases (10 per category), source images, and target transformations. The slash separates the textual components (left) from the visual components (right).

Text / Visual CueASR (%)HSEV (%)HRR (%)
r/(a,s) (SIJA)573.155554
(r,a)/s512.925048
(r,s)/a622.515438
a/(r,s)502.834846
s/(r,a)612.294632
(a,s)/r652.895847

The SIJA allocation achieves the highest HS (3.15) and HRR (54%), indicating the strongest production of valid high-risk edits among these allocations. The reversed (a,s)/r allocation yields higher ASR (65%) and EV (58%) but lower HRR (47%). This distinction is important: SIJA's asymmetric design favors harmful edit execution rather than simply maximizing refusal bypass. The information-preserving controls complement the component-removal analysis by directly testing the placement of intent components. (Revision: Sec. 5.3 and Appendix C.4.)

A2. Benign Controls and Safety–Utility Trade-off

We thank the reviewer for highlighting the importance of benign editing utility. We construct 100 benign split-intent controls (10 per category), preserving the source images, referent-identifying instructions, and graphical cue format while replacing unsafe transformations with benign edits. We evaluate benign acceptance under both the original conservative safety criterion and a relaxed criterion, while keeping each defense's intent-reconstruction procedure unchanged.

DefenseSafety SettingUnsafe Rejection (%)Benign Pass (%)
VJA IntrospectionOriginal68.444
Decomposed Introspection (Ours)Original93.111
VJA IntrospectionRelaxed48.755
Decomposed Introspection (Ours)Relaxed66.860

Unsafe rejection is evaluated on all 766 SI2ESBench cases, while benign acceptance is measured on the 100 controls. The original conservative criterion leads to substantial over-refusal, with our defense accepting only 11% of benign requests. Relaxing the safety criterion increases benign acceptance to 60%, while unsafe rejection changes from 93.1% to 66.8%. Notably, under the relaxed setting, our defense achieves both higher unsafe rejection (66.8% vs. 48.7%) and higher benign acceptance (60% vs. 55%) than VJA Introspection. These findings distinguish the contribution of editing-intent reconstruction from the conservativeness of the subsequent safety decision. Our central contribution is a defense principle that reconstructs the complete transformation from distributed multimodal inputs, rather than an optimized safety decision criterion. The results demonstrate its effectiveness for detecting split-intent attacks while highlighting the importance of safety calibration. Further improving this trade-off and evaluating end-to-end benign editing success remain directions for future work. (Revision: Sec. 5.4 and Appendix C.5.)

A3. Separating Intent Allocation from Visual Encoding

We appreciate the request to distinguish SIJA's contribution from graphical encoding already considered in VJA. We conduct a matched 2×22\times2 experiment using GPT Image 2 on 100 category-stratified cases sampled from SI2ESBench (10 per risk category), independently varying intent allocation and visual encoding while preserving the source images and intended transformations. Concentrated places the complete (r,a,s) intent in a visual input; Split places r in text and (a,s) in the visual cue. Textual uses readable text in the visual input, while Graphical uses non-textual graphical semantics.

Intent AllocationVisual EncodingASR (%)HSEV (%)HRR (%)
ConcentratedTextual422.063027
ConcentratedGraphical611.37188
SplitTextual472.163729
SplitGraphical (SIJA)573.155554

Graphical encoding lowers HRR from 27% to 8% when intent is concentrated, but raises HRR from 29% to 54% when intent is split. Conversely, splitting raises HRR by 46 percentage points with graphical cues (8% to 54%), compared with 2 points when visual instructions are textual (27% to 29%). This interaction isolates the key mechanism of SIJA: the effectiveness of graphical action/specification semantics increases substantially when the edit referent is independently grounded in text. (Revision: Sec. 5.3 and Appendix C.4.)

A4. Benefit of Explicit Decomposed Introspection

We thank the reviewer for pointing out Sun et al., VPA-Guard: Defending and Benchmarking Image-to-Video Generation Against Visual Prompt Attacks (2026). VPA-Guard uses retrieval-augmented few-shot safety prompting and jointly considers visual and textual cues for image-to-video generation. We now cite and discuss this related work (Secs. 2 and 3.2). Our defense focuses on source-conditioned image editing and tests the additional value of explicitly representing the inferred transformation as referent, action, and specification, beyond generic joint-intent reasoning. To measure that value, we introduce a Generic Joint-Intent Reconstruction baseline that infers the complete edit without a three-field decomposition. The generic and decomposed variants use matched safety criteria and inference settings. We evaluate them on the same 766 SI2ESBench cases and report the pre-generation safety inference cost.

DefenseRejection (%)Input / Output TokensLatency (s)
VJA Introspection68.41948.46 / 166.845.70
Generic Joint-Intent Reconstruction86.32069.46 / 117.474.02
Decomposed Introspection (Ours)93.12148.46 / 123.774.37

Generic reconstruction improves rejection by 17.9 percentage points over VJA Introspection. Explicit three-field decomposition increases rejection by a further 6.8 points over generic reconstruction, with an average increase of 6.30 output tokens and 0.35 seconds per request. These controlled results support the contribution of explicit intent decomposition to unsafe-edit detection, while the VPA-Guard comparison situates our work within multimodal safety reasoning. (Revision: Secs. 2, 3.2, and 5.4; Appendix C.5.)

A5. Instance-Level Judge Agreement

We thank the reviewer for highlighting that aggregate agreement need not imply sample-level consistency. We therefore evaluate the primary judge, Gemma-4-31B-it, against human consensus on the same 350 edited outputs used in the original analysis. Three annotators assess each output. Consensus uses majority voting for EV and median ratings for HS; HRR is positive when consensus EV is 1 and HS is at least 4.

MetricPrecision (%)Recall (%)F1 (%)Judge–Human κHuman–Human κ
EV94.2495.9795.100.7700.826
HRR95.1195.4795.290.8050.933
HS———0.8410.948

For EV and HRR, κ denotes Cohen's kappa; for ordinal HS, it denotes quadratic weighted kappa. The paired confusion matrices for the primary judge are:

MetricTPFPFNTN
EV262161161
HRR253131272

Gemma-4-31B-it achieves F1 scores of 95.10% for EV and 95.29% for HRR, with judge–human kappa of 0.770 and 0.805, respectively, and HS weighted kappa of 0.841. The confusion matrices directly characterize sample-level errors, while pairwise human agreement provides an annotation-reliability reference. We also report results for the other MLLM judges in the appendix. (Revision: Sec. 5.5 and Appendix A.4.)

Thank you again for your detailed and constructive feedback, which has motivated several important additional experiments.

We would like to briefly summarize how we have addressed your five main concerns:

  1. Information-preserving controls: We evaluated all six modality allocations while preserving the complete editing intent. SIJA achieves the highest HRR (54%), supporting the effectiveness of its asymmetric design.
  2. Benign utility: We introduced 100 benign controls and evaluated both original and relaxed safety criteria. Under the relaxed setting, our defense achieves higher unsafe rejection (66.8% vs. 48.7%) and benign acceptance (60% vs. 55%) than VJA Introspection. We also acknowledge the over-refusal under the original setting.
  3. Contribution beyond VJA: Our controlled 2×2 experiment independently varies intent allocation and visual encoding. The results reveal a strong interaction between split intent and graphical encoding, with HRR increasing from 8% to 54% under graphical cues when moving from concentrated to split intent.
  4. Explicit decomposition: Under matched safety criteria and inference settings, decomposed introspection improves unsafe rejection from 86.3% to 93.1% over generic joint-intent reconstruction.
  5. Judge reliability: We added instance-level agreement analysis against human annotations, obtaining F1 scores of 95.10% for EV and 95.29% for HRR, together with detailed agreement statistics and confusion matrices.

These additional results and the corresponding discussions have been incorporated into the revised manuscript. We hope they adequately address your concerns and provide stronger evidence for the paper's contributions, while also clarifying the remaining limitations.

Thank you again for your time and consideration.

Thank you again for your detailed and insightful feedback. We would like to further highlight the central novelty of our work and how the additional experiments strengthen its empirical foundation.

Our main contribution is identifying and systematically characterizing a split-intent safety vulnerability that cannot be explained by visual encoding alone. Unlike prior image-editing jailbreaks that concentrate the complete harmful intent in a single input component, SIJA distributes complementary editing semantics across modalities while preserving their joint recoverability. The information-preserving allocation experiments and controlled 2×2 study provide direct evidence that split-intent allocation and graphical encoding interact to enable harmful edit execution. Beyond attack effectiveness, SI2ESBench offers a structured evaluation setting for investigating this mechanism, while our editing-intent reconstruction defense demonstrates the benefit of explicitly recovering the referent, action, and specification before safety assessment. The additional benign-utility experiments and instance-level judge agreement analyses further clarify the defense's limitations and strengthen the reliability of our findings.

We hope these results collectively establish the distinct contribution of split-intent modeling, controlled evaluation, and transformation-level safety assessment. Thank you again for your careful consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.