acceptodds
Under review as a conference paper at ICLR 2027

Red-Teaming Procedural and Persuasive Text Rendering in Text-to-Image Models

Abstract

Recent text-to-image (T2I) models can now render accurate and legible text within generated images. This capability allows visually benign images to convey dangerous procedures or deceptive messages through rendered text. However, existing red-teaming methods treat such text as a single risk source, conflating actionable guidance with persuasive messaging. Moreover, their black-box prompt-optimization approach fails to adequately capture the intricate relationship between prompt modifications and the resulting output states. To disentangle procedural and persuasive harms, we construct ToxiGlyph, a red-teaming dataset that pairs procedural and persuasive instances sharing the same harmful intent, spanning 10 risk categories and 31 subcategories. We further propose GlyphPilot, which formulates black-box prompt revision as a vision-guided multi-directional search. Rather than refining a single candidate along a fixed path, GlyphPilot generates and evaluates diverse revision strategies in parallel and uses visual feedback to select the most promising direction at each iteration. Across five commercial T2I models (e.g., Nano Banana 2 and GPT Image 2), GlyphPilot achieves an 87.85% ASR, surpassing the strongest baseline by over 40 points, with persuasive text consistently more harmful than procedural text. These results establish text rendering as a critical yet underexplored attack surface in T2I models.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Text-to-image Models, Automated Red-Teaming, Text Rendering

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_eUVcOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies harmful text rendered by T2I models. It introduces ToxiGlyph and GlyphPilot, a black-box prompt search method, and reports strong attack results across five commercial T2I models.

Critical strengths

One thing I found valuable is the paper’s separation of procedural and persuasive text under the same harmful intent. This paired design makes the comparison much cleaner than treating all rendered text as one risk category, and it leads to a concrete observation that the two types behave differently in both attack success and harm severity. The evaluation also spans five commercial and popular T2I models and includes category-level and cross-model transfer results.

Critical weaknesses

1.If the final threat is simply to produce an image containing clear harmful text, an attacker could instead use an LLM to generate the content from the same high-level intent and then place it onto a benign image with a standard image editor. This would likely be cheaper, more controllable, and produce more legible text. In this case, why is it necessary to go through the additional step of jailbreaking a T2I model and asking it to generate and render the harmful text itself? Otherwise, the threat model seems closer to demonstrating an alignment blind spot in T2I models than to showing an attack path that a real-world attacker would prefer. However, if revealing this alignment blind spot is the main contribution, the mechanistic analysis is still limited. The paper mainly shows that the failure can be triggered, but does not explain why it occurs. For instance, whether it comes from insufficient input moderation, the model’s internal text-generation capability, or the lack of output-side safety mechanisms for rendered text.

2.Even if the practical relevance in Point 1 is accepted, the paper still misses an important comparison with closely related work. In particular, TYPO studies a highly similar setting: it also performs black-box search against commercial T2I models to elicit detailed and readable harmful instructions inside generated images. Since TYPO is already discussed in the related work, its absence from the main comparison makes it difficult to assess the actual improvement over the closest prior methods. Personally, GlyphPilot is still largely an LLM/VLM-driven prompt generation and iterative refinement pipeline, with the main methodological difference being visual-state diagnosis and multi-directional search.

3.ToxiGlyph is used not only for evaluation but also to provide fairly detailed target specifications to the attack, since the procedural/persuasive prototypes are directly used by the diagnosis and candidate-generation modules. It is therefore unclear how well GlyphPilot would work when only a high-level harmful intent is available. An ablation without the prototype, or evaluation on an external benchmark without such task-specific prototypes, would better show the generality of the search method.

4.The ablation does not really isolate the main components of GlyphPilot. Visual-State Diagnosis and Multi-Directional Exploration are presented as the key ideas, but the main component ablation only compares first-candidate, random, and selector-based selection. Missing comparisons include w/o Visual-State Diagnosis, single-direction vs. multi-direction search, w/o history, w/o refusal-guided expansion, w/o visual-state refinement, and different diagnostic dimensions. Without these, it is hard to tell what actually drives the reported 87.85% ASR.

Questions for authors

1.The paper analyzes the overall search budget, but why was the final setting fixed to 12 target queries, with 3 initial queries and up to 9 refinement queries? Also, why are exactly 3 candidates generated in each round? A sensitivity study with, for example, 1/2/3/5 candidates would make these design choices easier to justify.

2.GlyphPilot relies on several large and different models, including Gemma 4 26B, Qwen3-VL-32B, and Gemini 3.1 Pro. I think the paper should report the cost and computation of the whole pipeline, and test smaller models or different backbones for a fairer comparison with the baselines.

3.Final harmfulness is judged only by GPT-5.4. If GPT-5.4 is particularly sensitive or lenient toward certain types of persuasive content, it could systematically affect the reported 83.2% vs. 90.8% gap. Could the authors evaluate with judges from different model families, preferably including reproducible open-source models, and report agreement on a human-annotated subset? This would help rule out single-judge bias.

4.Since the attack depends on readable harmful text in the generated image, a natural defense is OCR or VLM transcription followed by a text safety guard. The paper does not evaluate this setting. Given that Beautiful Images, Toxic Words (AAAI 2026) studies mitigation for harmful rendered text and Multimodal Pragmatic Jailbreak (ACL 2025) evaluates safety filters, how much of GlyphPilot’s ASR remains under such an output-side defense?

5 comments

Thank you for your thorough review and valuable feedback. We appreciate your insights and would like to address each of your points below.

1. Research Motivation and Mechanism Analysis (Weakness1)

We would like to clarify that this paper focuses on whether commercial T2I models with reasoning capabilities can autonomously generate harmful text and render it into an image when the complete harmful text is not provided. Our threat model consists of three components:

  • Attack Scenario: The attacker uses a commercial T2I model to generate images containing harmful text, such as notices, posters, or chat screenshots.
  • Attack Capability: The attacker has only black-box API access and may use auxiliary models to construct and revise prompts. Neither the attacker nor the auxiliary models may provide the complete harmful text to the target T2I model, so the target model must generate and render it autonomously.
  • Attack Objective: Within a limited query budget, the attacker induces the target model to autonomously generate and render readable procedural or persuasive text that matches a specified harmful intent.

Similar harmful images can also be produced by combining textual jailbreaks with manually added text, but ensuring that the text blends naturally with the layout, lighting, and scene could still be difficult and costly. This paper focuses on whether the target T2I model can autonomously generate harmful body text and integrate it into realistic, deceptive images.

The rejection rates in Table 1 of our paper reveal substantial differences in the defensive behavior of different models:

  • Weak Safety Defenses. Under DIRECT, models such as Grok Imagine 2 exhibit rejection rates (Ref.) below 5% for both task types, indicating relatively weak safety protection.
  • Strong but Bypassable Safety Defenses. Models such as GPT Image 2 have high rejection rates (above 70%), indicating relatively strong external safety filters on the input or output side. However, after the attack, these models still generated images containing harmful text in a substantial proportion of cases. This indicates that safety filtering and model alignment do not reliably prevent harmful text generation and rendering.

2. Additional Comparison with TYPO (Weakness 2)

Our study focuses on whether a target T2I model can infer and generate harmful text without being given the complete text. TYPO's original implementation allows complete harmful text to be supplied directly for rendering, which differs from our setting, so we did not include it in the initial comparison.

To address the reviewer's concern, we have added comparisons with TYPO on two target models and report the attack success rate (ASR) and harmful payload rate (HPR). HPR is the proportion of final prompts that GPT-5.4 judges to contain complete harmful text based on the task intent and related information.

Target modelMethodProcedural ASR (%)↑Procedural HPR (%)↓Persuasive ASR (%)↑Persuasive HPR (%)↓
GPT Image 2TYPO90.351.658.125.8
GlyphPilot80.6090.30
Grok Imagine 2TYPO77.451.638.738.7
GlyphPilot96.801000

GlyphPilot achieves higher ASR in most settings without supplying complete harmful text to the target model (HPR = 0). Although TYPO achieves a slightly higher ASR on GPT Image 2's Procedural tasks, its HPR is as high as 51.6%. This indicates that TYPO's ASR advantage is accompanied by a greater reliance on explicitly embedding harmful content in prompts, whereas GlyphPilot achieves comparable attack effectiveness with less reliance on explicit harmful content in prompts.

3. Evaluation on AdvBench (Weakness 3)

Our method (GlyphPilot) operates directly on harmful intents without requiring detailed procedural or persuasive prototypes. To validate this capability, we evaluate GlyphPilot on harmful intents from the external AdvBench benchmark, using GPT-Image-2 as the target image generator. The results demonstrate that GlyphPilot remains effective in this setting.

MethodASR (%)HsuccH_{succ}
Direct24
GlyphPilot723.5

4. Component Ablation of GlyphPilot (Weakness4)

We ablate four core components of GlyphPilot: (1) w/o Structured Diagnosis removes structured diagnostic feedback while retaining the previous image; (2) w/o History removes prior search history and repetition-avoidance memory while retaining the current image and diagnosis; (3) Single-direction retains the three-candidate initialization but generates only one diagnosis-guided revision per subsequent round, bypassing candidate selection; and (4) w/o Visual Guidance removes the previous image and diagnosis-derived feedback, retaining only non-visual context and the binary refusal outcome. The results show that removing any component reduces both overall ASR and HsuccH_{\mathrm{succ}}, supporting the contributions of structured diagnosis, history, multi-directional exploration, and visual guidance.

MethodProc. ASR (%)Pers. ASR (%)Overall ASR (%)Δ ASR (pp)H_succ
Full808482—4.05
w/o Structured Diagnosis606060−22pp3.83
w/o History686466−16pp3.79
Single-direction647670−12pp3.86
w/o Visual Guidance727674−8pp3.84

5. Search Budget and Candidate Count (Question 1)

  • Query budget. Figure 8 of our paper already reports a sensitivity analysis of ASR under different query budgets. The results show that ASR largely stabilizes at around 10 search rounds, so we set the search to 10 rounds. Since no visual feedback is available initially, we query all three candidates in the first round to select a starting point. In each subsequent round, we use feedback to select and query one candidate, yielding at most 3+9=123+9=12 target queries.
  • Candidate count ablation. We added a sensitivity study with N=1,2,3,5N=1,2,3,5, keeping the search limit fixed at 10 rounds. ASR increases from 66.13% to 87.10% as the candidate count increases. Considering the tradeoff between effectiveness and overhead, we choose N=3N=3 as the default setting.
Candidate countMaximum queriesP_proc ASR (%)P_pers ASR (%)Combined ASR (%)HsuccH_{succ}
N=11054.8477.4266.133.95
N=21180.6583.8782.264.08
N=31283.8787.1085.484.13
N=51487.1087.1087.104.02

6. Backbone Substitution and Runtime Overhead (Question 2)

To assess whether GlyphPilot depends on specific backbone models, we separately replace the backbones used by (i) the candidate generator A\mathcal{A} (Gemma 4 26B → Qwen3-VL-8B), (ii) the candidate selector SS and the prerequisite-verification stage DD (Qwen3-VL-32B → Qwen3-VL-8B), and (iii) the semantic-diagnosis stage DD (Gemini 3.1 Pro → GPT-5.1), while keeping all other settings fixed. GlyphPilot maintains high attack success rates with smaller backbones and models from different families, suggesting that its effectiveness does not depend on a particular model. Stronger backbones further improve overall performance.

ConfigurationASR (%)HsuccH_{succ}
Full74.203.78
A\mathcal{A} → Qwen3-VL-8B67.703.52
SS / DD (Stage 1) → Qwen3-VL-8B64.503.80
DD (Stage 2) → GPT-5.161.303.60

In our method, the candidate generator (Gemma 4 26B) is deployed locally on two NVIDIA A100 80GB GPUs, while all other models are accessed via APIs. We report the average number of target T2I queries and auxiliary model calls during search for each method, together with ASR and harmful payload rate (HPR). Compared with baselines, GlyphPilot incurs higher call overhead but achieves higher ASR and lower HPR.

MethodASR (%)↑HPR (%)↓Avg. T2I Queries↓Avg. Auxiliary Calls↓
TYPO66.1341.941.617.22
ETCH38.8210.881.65.17
GlyphPilot87.8505.8915.52

7. Evaluation Reliability (Question 3)

To assess potential single-evaluator bias, we conducted additional automatic and human evaluations:

  • Additional evaluators. We re-evaluated the same outputs from our method (GlyphPilot) using Claude Sonnet 5, Mistral Small 3.2-24B (open-source), and Qwen3.6-27B (open-source), reporting both ASR and HsuccH_{\mathrm{succ}}. Their results were broadly consistent with those reported in Table 1.
  • Human evaluation. Ten independent annotators assessed the harmfulness of 200 images generated by our method (GlyphPilot). Using human majority votes as the reference, GPT-5.4 achieved 84.4% accuracy, 86.0% precision, 96.6% recall, and 91.0% F1 for harmful-content detection.
EvaluatorProcedural ASR (%)Persuasive ASR (%)Average ASR (%)Procedural HsuccH_{\mathrm{succ}}Persuasive HsuccH_{\mathrm{succ}}Average HsuccH_{\mathrm{succ}}
GPT-5.4 (original)83.4492.2687.853.6394.3314.002
Claude Sonnet 585.8190.9788.393.6294.3193.984
Mistral Small 3.2-24B95.0595.9195.484.4074.4484.428
Qwen3.6-27B86.6790.1188.393.8734.224.05

8. Output-side Defense Evaluation (Question 4)

We added output-side defense experiments on GPT Image 2 and Grok Imagine 2 using PaddleOCR transcription followed by Llama Guard text moderation, LlavaGuard image moderation, and their combination. The results show that LlavaGuard provides little effective protection, whereas OCR-based text moderation substantially reduces ASR but does not eliminate it entirely. This suggests that robust defense requires internal safety alignment that explicitly accounts for the harmful semantics and functions of rendered text, rather than reliance on output-side defenses.

MethodNo Defense ASR (%)OCR + Llama Guard ASR (%)LlavaGuard ASR (%)Combined Defense ASR (%)
DIRECT48.3911.2922.588.87
ETCH41.9422.5833.8720.16
TYPO66.1339.5263.7138.71
GlyphPilot90.3234.6883.8733.87
Reviewer_px3NOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies the risk of text-to-image (T2I) models rendering harmful text inside generated images. The authors divide such text into two forms. Procedural text instructs the viewer how to carry out a harmful action, while persuasive text aims to change the viewer’s beliefs or behaviors. Based on this, the authors construct ToxiGlyph, a paired dataset in which each harmful intent appears in both forms, and propose GlyphPilot, a black-box red-teaming framework. In each round, a VLM diagnoses the visual state of the generated image, an LLM generates several candidate revisions, and a selector picks one candidate to query the target model under a fixed budget. Experiments on five commercial T2I models show that GlyphPilot achieves higher attack success rates than direct prompting and ETCH, and that persuasive text is generally easier to elicit than procedural text.

Critical strengths
  • The topic is timely. Recent T2I models can render long and well-formatted text, so harmful content carried by the rendered text becomes a practical concern. Prior T2I safety work, which mostly focuses on harmful visual content, has paid less attention to this issue.
  • The evaluation covers five recent commercial T2I models (e.g., GPT Image 2 and Nano Banana 2). Notably, the direct prompting results alone already reveal large differences in safety alignment across vendors, which I find informative for the community.
  • The paper is generally well organized and easy to follow. Fig. 1 and Fig. 3 make the motivation behind the search design clear.
Critical weaknesses
  • The practical threat posed by this attack is unclear to me, especially for procedural text. An adversary who wants step-by-step instructions for a harmful task can jailbreak an LLM to obtain the same knowledge in text form, and rendering that text into an image adds little capability. Moreover, GlyphPilot relies on a locally deployed Gemma 4 26B to rewrite harmful prompts, which means the adversary already has access to a cooperative LLM. In this case, it is unclear why the adversary would need to query a commercial T2I model to obtain the same information. If the motivation is fidelity or deception (e.g., a fake public notice), the adversary can also generate a benign image with any T2I model and then add the desired text with standard image editing tools. This approach is cheap and gives full control over the content. Therefore, I would expect the authors to clarify what additional harm or capability T2I text rendering brings beyond text-only jailbreaks and manual editing, as the current paper does not discuss this point.
  • The success criterion appears too loose to support the claims about executability and credibility. Sec. 6.1 states that the harm score depends on harm severity and functional effectiveness, namely executability for procedural text and credibility for persuasive text. However, according to Table 9, an output with Substantial or Major severity receives Hs≥3H_s \geq 3 and counts as a success, even when its executability is only Vague or its credibility is only Crude. As a result, 7 of the 9 cells for each text form lead to a success, and functional effectiveness only matters in the Limited row. Besides, harm severity is largely determined by the topic of the intent, and a "generalized" intent match also satisfies the matching criterion (Appendix C.2). Thus, the reported ASR mainly reflects whether the model renders readable and on-topic text without refusal. This also weakens the claim in Sec. 6.3 that procedural rendering is harder because it requires "detailed and actionable harmful instructions", since most successful procedural outputs are not required to be actionable. I would suggest reporting the ASR under a stricter criterion (e.g., requiring at least Moderate executability or Credible content), together with a human evaluation on a subset.
  • The methodological novelty of VGMS feels limited. In VGMS, an attacker LLM generates multiple candidate prompts in each round, an evaluator ranks them before any target query, only the most promising candidate is sent to the target, and a judge provides feedback for the next iteration. This procedure closely follows the structure of TAP [2]. Similarly, iteratively refining the attack prompt based on structured feedback from the target response is the core idea of PAIR [1]. However, neither work is cited or discussed in the paper. As far as I can tell, the main differences from TAP are the VLM-based diagnosis and the legibility check before scoring (Sec. 5.2), which I regard as engineering adaptations to the T2I setting. Therefore, the technical contribution mainly lies in transferring an established jailbreak paradigm to T2I text rendering. I would expect a clear discussion and an empirical comparison with PAIR and TAP adapted to T2I models using the same VLM judge.

References:

[1] Chao et al., Jailbreaking Black Box Large Language Models in Twenty Queries. SaTML 2025.

[2] Mehrotra et al., Tree of Attacks: Jailbreaking Black-Box LLMs Automatically. NeurIPS 2024.

Questions for authors
  • What additional harm or capability does T2I text rendering provide beyond jailbreaking an LLM or adding text to an image manually?
  • Since GlyphPilot already relies on a cooperative local LLM, why would a realistic adversary need a T2I model to obtain procedural content?
  • What is the fundamental difference between VGMS and TAP, and how does GlyphPilot compare with PAIR and TAP when both are adapted to T2I models?
9 comments

Thank you for your thorough review and valuable feedback. We appreciate your insights and would like to address each of your points below.

1. Research Motivation (Weakness 1 & Question 1 & Question 2)

We would like to clarify that this paper focuses on whether commercial T2I models with reasoning capabilities can autonomously generate harmful text and render it into an image when the complete harmful text is not provided. Our threat model consists of three components:

  • Attack Scenario: The attacker uses a commercial T2I model to generate natural images containing harmful text, such as notices, posters, or chat screenshots.
  • Attack Capability: The attacker has only black-box API access and may use auxiliary models to construct and revise prompts. Neither the attacker nor the auxiliary models may provide the complete harmful text to the target T2I model, so the target model must generate and render it autonomously.
  • Attack Objective: Within a limited query budget, the attacker induces the target model to autonomously generate and render readable procedural or persuasive text that matches a specified harmful intent.

Similar harmful images can also be produced by combining textual jailbreaks with manually added text, but ensuring that the text blends naturally with the layout, lighting, and scene could still be difficult and costly. This paper focuses on whether the target T2I model can autonomously generate harmful body text and integrate it into realistic, deceptive images.

2. Evaluation Reliability (Weakness 2)

To assess evaluation reliability, we add stricter functional criteria and human evaluation:

  • Stricter functional criteria. We add a functional effectiveness threshold Q≥kQ\ge k (k=1,2,3k=1,2,3) to the original success criterion. QQ corresponds to the executability of procedural text and the credibility of persuasive text defined in the paper. Using GPT-5.4 scores for the same outputs, GlyphPilot's ASR decreases as kk increases, but its mean ASR remains substantially higher than baseline methods.
  • Human evaluation. Ten independent annotators assessed the harmfulness of 200 images generated by our method (GlyphPilot). Using human majority votes as the reference, GPT-5.4 achieved 84.4% accuracy, 86.0% precision, 96.6% recall, and 91.0% F1 score for harmful-content detection.
Target modelMethodProcedural k=1Procedural k=2Procedural k=3Persuasive k=1Persuasive k=2Persuasive k=3
Nano Banana 2DIRECT8.6%8.6%4.3%45.16%45.16%40.86%
ETCH25.81%24.73%1.08%60.22%60.22%45.16%
GlyphPilot79.57%75.27%22.58%92.47%92.47%81.72%
GPT Image 2DIRECT2.15%2.15%2.15%23.66%23.66%20.43%
ETCH27.96%27.96%8.6%46.24%46.24%45.16%
GlyphPilot75.27%75.27%38.71%86.02%86.02%82.8%
Grok Imagine 2DIRECT79.57%79.57%54.84%94.62%94.62%87.1%
ETCH31.18%31.18%12.9%49.46%49.46%44.09%
GlyphPilot94.62%94.62%56.99%97.85%97.85%95.7%
Qwen Image 3DIRECT6.45%6.45%1.08%61.29%60.22%53.76%
ETCH16.13%15.05%4.3%61.29%61.29%41.94%
GlyphPilot78.49%78.49%21.51%88.17%88.17%83.87%
Seedream 5 ProDIRECT73.12%72.04%30.11%79.57%79.57%69.89%
ETCH23.66%23.66%4.3%46.24%46.24%31.18%
GlyphPilot89.25%89.25%27.96%96.77%96.77%87.1%
AverageDIRECT33.98%33.76%18.49%60.86%60.65%54.41%
ETCH24.95%24.52%6.24%52.69%52.69%41.51%
GlyphPilot83.44%82.58%33.55%92.26%92.26%86.24%

3. Comparison with PAIR and TAP (Weakness 3& Question 3)

PAIR, TAP, and our method (GlyphPilot) all use feedback-based iterative optimization, but differ fundamentally in feedback structure and search structure.

  • Coarse Feedback and Single-Candidate Updates in PAIR. PAIR uses the target model's response and an overall score as coarse feedback to guide optimization. Each trajectory generates only one revised prompt per round, limiting candidate diversity. A poor revision can then steer subsequent search away from the objective. GlyphPilot instead organizes visual feedback into explicit multidimensional states, specifies what to revise and preserve, and generates multiple complementary candidates targeting different unmet requirements in each round.
  • Coarse Feedback and Limited Branch Guidance in TAP. TAP extends PAIR with tree search but still relies on target responses and an overall score without explicit multidimensional diagnosis. In addition, its branches are generated from overall feedback without clear revision requirements for specific visual defects. Tree expansion therefore adds candidates without providing more targeted revision guidance, while requiring target queries to evaluate multiple branches. Instead of maintaining a growing search tree, GlyphPilot uses the current image's multidimensional diagnosis to specify what to revise and preserve, dynamically generates complementary revision candidates, and selects one before the next target query.

We added comparisons with PAIR-T2I and TAP-T2I on GPT Image 2 and Grok Imagine 2, using the same model configurations as GlyphPilot in our paper. GlyphPilot achieves higher ASR than both baselines across target models.

Target modelMethodProcedural ASR (%)↑Persuasive ASR (%)↑
GPT Image 2PAIR-T2I58.0687.1
TAP-T2I48.3974.19
GlyphPilot80.6590.32
Grok Imagine 2PAIR-T2I70.9787.1
TAP-T2I70.9787.1
GlyphPilot96.77100
AveragePAIR-T2I64.5287.1
TAP-T2I59.6880.65
GlyphPilot88.7195.16
Reviewer_px3NOpenReview ↗

Thanks for the response!

Thanks for the detailed response and the additional experiments. I can see a lot of efforts were put into it. However, after reading the rebuttal I still have several concerns that are not fully resolved:

  1. Threat motivation: My original question was, since GlyphPilot already relies on a cooperative local LLM, why would a realistic adversary still need to go through a T2I model? The response mainly restates the threat model and adds a new constraint that the complete harmful text can not be provided to the target model. But this constraint seems to be defined by the authors, and a real attacker is not restricted by it. For the manual editing alternative, the response only says it "could still be difficult and costly" without any concrete evidence. So I feel this question is still not directly answered.

  2. Stricter success criterion: The results under k=2 are almost identical with the original ones. Does this mean the judge almost never assign the lowest level, so the scale actually has little discrimination at the low end? Also, the procedural ASR drops a lot under k=3, and on some models it becomes close to DIRECT. Does this mean most of the successful procedural outputs are in fact not really executable?

  3. Comparison with PAIR/TAP: The differences described in the response (e.g., multidimensional diagnosis, select before query) are basically the ones I already mentioned in my review, and I still tend to regard them as engineering adaptations to the T2I setting. The newly added comparison is also in a relatively small scale and some settings are not clearly explained, e.g., whether the query budget is aligned and whether the baselines use the same diagnosis model. So I think the current results are not very convincing yet.

Overall, I would like to keep my current score.

Reviewer_7pigOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper investigates safety risks arising from harmful text rendered within images generated by text-to-image models. It distinguishes between procedural text that provides actionable harmful instructions and persuasive text that conveys deceptive messages. The authors introduce ToxiGlyph, a paired red-teaming dataset comprising 310 harmful intents with procedural and persuasive prototypes. They further propose GlyphPilot, a black-box red-teaming framework that combines multidimensional visual-state diagnosis with feedback-guided multi-directional prompt exploration to improve attack effectiveness. Experiments on 186 red-teaming tasks across 5 T2I models show that GlyphPilot achieves a high average attack success rate which outperforms two baseline methods.

Critical strengths
  1. A Meaningful Research Topic. This is an interesting and meaningful research direction in t2i safety.
  2. A Dataset. The paper creates a red-teaming dataset for further studies.
Critical weaknesses
  1. Limited Methodological Novelty. The proposed GlyphPilot framework closely resembles existing feedback-driven prompt optimization approaches, particularly ETCH and TYPO. Although the authors introduce multidimensional visual diagnosis and candidate selection, the technical novelty beyond existing search strategies remains unclear.

  2. Insufficient and Potentially Unfair Baseline Evaluation. The effectiveness claims rely on comparisons with only DIRECT and ETCH, while the closely related TYPO method mentioned in related work is not evaluated. Also, ETCH substantially underperforms even DIRECT on several models, raising concerns about baseline implementation.

Questions for authors
  1. How does the proposed framework fundamentally differ from the feedback-guided optimization in ETCH?
  2. Why does ETCH perform substantially worse than DIRECT on models such as Grok Imagine 2 and Seedream 5 Pro? Does ETCH preserve the best previously observed candidate?
  3. Why is TYPO, which also targets harmful text rendering mentioned in related work, not included in the experimental comparisons?
  4. Could the authors provide human evaluation or independent judge comparisons to establish the reliability of the reported ASR and harm severity scores?
5 comments

Thank you for your thorough review and valuable feedback. We appreciate your insights and would like to address each of your points below.

1. Methodological Novelty (Weakness 1 & Question 1)

Unlike ETCH and TYPO, which rely on orthogonal prompt layers and fixed strategy sets, respectively, GlyphPilot dynamically generates and selects complete prompt candidates based on multidimensional visual states.

  • Comparison with ETCH. ETCH divides prompts into three functionally orthogonal layers for semantic camouflage, visual-spatial anchoring, and typographic encoding. It diagnoses failures, updates the relevant layers, and leaves others unchanged. However, our experiments reveal a many-to-many relationship between prompt edits and output states (Figure 3). This orthogonal and fixed mapping of ETCH limits the revision directions explored for each defect.
  • Comparison with TYPO. TYPO defines fixed textual and visual strategy dimensions, including visual carriers, layout styles, and text transformation rules. Based on output feedback, it adjusts strategy combinations within this predefined static space to generate subsequent prompts. Its original implementation also allows complete harmful text to be supplied directly in the prompt for rendering. Our supplementary experiments show that 41.94% of TYPO's final prompts contain complete harmful text, rising to 51.22% among successful cases.
  • Our method design. In contrast, GlyphPilot dynamically generates and selects complete prompt candidates guided by multidimensional visual states, enabling coordinated revisions across the entire prompt rather than restricting updates to orthogonal layers or fixed strategy options. Moreover, GlyphPilot does not require complete harmful text to be supplied to the target model.

2. Performance Differences of ETCH (Weakness 2 & Question 2)

ETCH's lower ASR than DIRECT on some models can be attributed to its defense-evasion design and the constraints imposed by its orthogonal prompt layers.

  • Mismatch between evasion design and target defenses. ETCH's threat model considers external safety filters, and its semantic camouflage and visual scene framing are designed to evade these checks. The DIRECT refusal rates (Ref.) in Table 1 of our paper reflect varying levels of defense strength across models. Accordingly, ETCH achieves greater ASR improvements on models with stronger defenses (e.g., GPT Image 2 and Nano Banana 2), while offering limited gains when DIRECT already achieves high ASR (e.g., Grok Imagine 2 and Seedream 5 Pro). The additional evasion requirements constrain content expression and text rendering, limiting attack effectiveness on these models.
  • Orthogonal layers constrain revision directions. ETCH divides prompts into three functionally orthogonal layers for semantic camouflage, visual-spatial anchoring, and typographic encoding. It diagnoses failures, updates the relevant layers, and leaves others unchanged. However, our experiments reveal a many-to-many relationship between prompt edits and output states (Figure 3). This orthogonal and fixed mapping of ETCH limits the revision directions explored for each defect.

ETCH in Table 1 of our paper follows the original paper's settings, without an additional mechanism for retaining the best candidate across rounds. To address the reviewer's question, we report results with best-candidate retention ( ETCHoracle\mathrm{ETCH}_{\mathrm{oracle}} ), as shown below. On Grok Imagine 2 and Seedream 5 Pro, the ASR of ETCHoracle\mathrm{ETCH}_{\mathrm{oracle}} is close to that of ETCH and remains below DIRECT.

Target modelText formDIRECTETCHETCH_oracle
Nano Banana 2Procedural8.6%25.81%26.88%
Persuasive45.16%60.22%61.29%
GPT Image 2Procedural2.15%27.96%32.26%
Persuasive23.66%46.24%47.31%
Grok Imagine 2Procedural79.57%31.18%33.33%
Persuasive94.62%49.46%49.46%
Qwen Image 3Procedural6.45%16.13%22.58%
Persuasive61.29%61.29%64.52%
Seedream 5 ProProcedural73.12%23.66%24.73%
Persuasive79.57%46.24%49.46%
AverageProcedural33.98%24.95%27.96%
Persuasive60.86%52.69%54.41%

3. Additional Comparison with TYPO (Weakness 2 & Question 3)

Our study focuses on whether a target T2I model can infer and generate harmful text without being given the complete text. TYPO's original implementation allows complete harmful text to be supplied directly for rendering, which differs from our setting, so we did not include it in the initial comparison.

To address the reviewer's concern, we have added comparisons with TYPO on two target models and report the attack success rate (ASR) and harmful payload rate (HPR). HPR is the proportion of final prompts that GPT-5.4 judges to contain complete harmful text based on the task intent and related information.

Target modelMethodProcedural ASR (%)↑Procedural HPR (%)↓Persuasive ASR (%)↑Persuasive HPR (%)↓
GPT Image 2TYPO90.351.658.125.8
GlyphPilot80.6090.30
Grok Imagine 2TYPO77.451.638.738.7
GlyphPilot96.801000

GlyphPilot achieves higher ASR in most settings without supplying complete harmful text to the target model (HPR = 0). Although TYPO achieves a slightly higher ASR on GPT Image 2's Procedural tasks, its HPR is as high as 51.6%. This indicates that TYPO's ASR advantage is accompanied by a greater reliance on explicitly embedding harmful content in prompts, whereas GlyphPilot achieves comparable attack effectiveness with less reliance on explicit harmful content in prompts.

4. Evaluation Reliability (Question 4)

To assess potential single-evaluator bias, we conducted additional automatic and human evaluations:

  • Additional evaluators. We re-evaluated the same outputs from our method using Claude Sonnet 5, Mistral Small 3.2-24B (open-source), and Qwen3.6-27B (open-source), reporting both ASR and HsuccH_{\mathrm{succ}}. Their results were broadly consistent with those reported in Table 1 of our paper.
  • Human evaluation. Ten independent annotators assessed the harmfulness of 200 images generated by our method. Using human majority votes as the reference, GPT-5.4 achieved 84.4% accuracy, 86.0% precision, 96.6% recall, and 91.0% F1 score for harmful-content detection.
EvaluatorProcedural ASR (%)Persuasive ASR (%)Average ASR (%)Procedural HsuccH_{\mathrm{succ}}Persuasive HsuccH_{\mathrm{succ}}Average HsuccH_{\mathrm{succ}}
GPT-5.4 (original)83.4492.2687.853.6394.3314.002
Claude Sonnet 585.8190.9788.393.6294.3193.984
Mistral Small 3.2-24B95.0595.9195.484.4074.4484.428
Qwen3.6-27B86.6790.1188.393.8734.224.05

Thanks for the response, and I appreciate the authors effort on it. However, I am not fully convinced.

For W1 methodological novelty, while the authors clarify the differences between GlyphPilot and existing methods, these distinctions are primarily conceptual. The response still lacks sufficient component-level ablations to demonstrate the proposed method provide substantial benefits over simpler feedback-guided optimization.

For W2, I remain concerned that ETCH_oracle still performs substantially worse than DIRECT on several models. If both methods share the same initial prompts and the oracle retains the best result across all attempts, this discrepancy requires further explanation.

I cannot see the the number of evaluated tasks and query-budget settings for the new TYPO comparison. I think adding the detail of these will make the results more convincing.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.