acceptodds
Withdrawn from ICLR 2027

The Rendering Gap: Where Should Moderation Look?

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Software agents increasingly ship images as code. A chart is a matplotlib script, a poster is an SVG, and such artifacts can travel through storage, mail filters, and moderation without ever existing as pixels. Yet the harm of this content often lives only in the rendered image: instructions fragmented into one glyph per element leave no harmful substring in the source, and symbols drawn from pure geometry carry no text at all. We show that where a moderation filter sits relative to rendering decides whether it sees such harm at all. On a preregistered benchmark of 770 stimuli spanning four code carriers and a five-level obfuscation gradient, every locally deployable guard we test, including the industry-standard LlamaGuard-3, detects at most 2.7% of the semantically invisible content on the code path (four of six guards exactly zero) while flagging nearly all of the same content shown as plain text. The failure is placement, not capability. Frontier API models can read obfuscated code, but the endpoints in front of them, input filters and quota limits, can block up to 69% of the calls that would carry it, so the problem moves one step upstream. The threat is cheap to realize: a thirty-word “rendering service” system role persuades three of four model families to write harmful payloads verbatim into SVG. Turning the diagnosis into a defense, Rasterize–Transcribe–Judge (RTJ) routes the artifact through rendering, asks a vision model to transcribe what it sees, and lets the text-guard class judge the transcription. RTJ is the only single path we test that detects at least 60% on every text condition in the benchmark, including the adversarial layouts where classical OCR collapses to 10%, at a false-positive cost of 1 out of 25. We release the benchmark, the method, and a placement decision framework with measured costs.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
co️ntent moderation, LLM safety, jailbreaks and obfuscation, vision-language models, safety benchmark

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_B6THOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies the "rendering gap" in content moderation for code-generating agents. When agents produce visual artifacts such as SVGs, inline HTML, or matplotlib scripts, text-based code scanners easily fail if the harmful content is fragmented per glyph or drawn via coordinates, even though the visual harm remains clearly legible to humans once rendered. The authors present RenderGapBench to evaluate local and frontier guards across both code and pixel representations, and propose Rasterize-Transcribe-Judge (RTJ)—a pipeline that renders the artifact, transcribes the visible text with a vision-language model, and delegates the safety judgment to a text guard.

Critical strengths
  1. The paper clearly articulates the "rendering gap" in code-generating agents, showing how simple text fragmentation or coordinate drawing in code can cleanly bypass standard text moderation while still delivering harmful visual content to the user.

  2. The underlying failure is demonstrated to be widespread rather than an isolated quirk, backed by extensive testing across both local open-weight guardrails and commercial frontier APIs.

  3. The constructed RenderGapBench benchmark is systematic and practical, thoughtfully testing multiple code carriers—such as SVG, inline HTML, and matplotlib—across varied levels of obfuscation.

  4. The proposed Rasterize-Transcribe-Judge (RTJ) pipeline is conceptually clean and empirically solid, demonstrating genuine effectiveness against text-based obfuscation and holding up much better under tricky layout shifts than traditional OCR or direct pixel-level baselines.

Critical weaknesses

Although this paper proposes the interesting concept of the "rendering gap" and addresses it with the designed RTJ approach, I still have several concerns.

  1. The core motivation behind the "rendering gap" feels somewhat disconnected from real security threats. While mainstream safety research on jailbreaks and adversarial attacks focuses on proactively inducing models to generate novel harmful knowledge or unintended behaviors, this paper mainly forces an agent into a passive repetition, simply typesetting the user's exact harmful text into an SVG. In practical scenarios, an adversary who already possesses the explicit harmful text can easily render it locally without querying an LLM, making this setup feel artificial and lacking a clear real-world threat justification.

  2. While RTJ demonstrates decent empirical defense against text obfuscation, it is essentially an incremental pipeline combining existing off-the-shelf components, including a headless renderer, VLM transcription, and a text guardrail, offering limited methodological novelty.

  3. RTJ completely fails against the pure-geometry threat that the paper itself heavily motivates, achieving an exact 0% detection rate on coordinate-drawn hate and violence symbols as shown in Table 2 and Table 5.

  4. The manuscript's presentation needs substantial improvement, containing noticeable uncompiled citation errors such as line 152 (?) and line 454 (?Jain et al., 2023) that reflect a lack of basic proofreading prior to submission.

  5. There is a direct contradiction between the main text and the appendix regarding experimental rigor: while Section 5.2 reports severe pixel-path collapses under layout perturbations as valid attacks, Appendix K quietly admits the preregistered human legibility test was never completed before submission and the numbers are reported unqualified. I have strong impression of a rushed auto-research pipeline without responsible human checking.

Questions for authors

See the weakness (1-5) above.

Reviewer_t5XYOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper points out a rendering gap where a potentially adversarial content stored as code, such as HTML, SVG, or matplotlib scripts, but presented visually to users. It shows that fragmenting or encoding harmful text can circumvent text-based guard models from recognizing its harmfulness in the source code. To mitigate such problems, the authors propose Rasterize-Transcribe-Judge (RTJ) piepline which renders the code into an image, transcribes visible text using a VLM, and applies a text safety guard model. The authors evaluate RTJ under several adversarial layout variations they introduce and report gains over OCR-based detection in some conditions, particularly when characters overlap.

Critical strengths

S1. The evaluation covers multiple open-weight models and commercial API models. The use of a broad range of guard models helps assess whether the rendering gap is consistently observed across different models.

Critical weaknesses

W1. Performance gains of RTJ are mainly seen in specific cases. RTJ substantially outperforms an OCR-based baseline (EasyOCR) when characters overlap. However, EasyOCR performs slightly better overall on the remaining text samples, including the naturalistic model outputs. It would be helpful to clarify when the additional VLM transcription step provides a meaningful benefit over OCR.

W2. The methodological novelty appears limited. Prior work [1] already uses OCR on webpage screenshots to address HTML obfuscation and passes the extracted text to an LLM to support phishing detection. Given that the gains over OCR are mainly seen in specific cases (as discussed in W1), it would be helpful to clarify what methodological contribution RTJ offers beyond replacing OCR with VLM-based transcription.

W3. The interpretation of API failures requires a clearer explanation. The quota-related API failures discussed in the caption of Table 3 reflect usage limits during the experiments rather than failures in frontier models’ ability to detect harmful content. Although the paper reports a rerun, the abstract groups these failures with input-filter rejections when reporting the API rejection rate. This framing may mislead readers to interpret quota-related failures as limitations of frontier models’ moderation capabilities. Since the authors have already reported the rerun results, the quota-related rejection rate from the earlier run could be omitted from the abstract to avoid this ambiguity.


[1] Liu et al., Less Defined Knowledge and More True Alarms: Reference-based Phishing Detection without a Pre-defined Reference List, USENIX Security 2024.

Questions for authors

Q1. Could the authors further clarify when RTJ outperforms EasyOCR? RTJ performs slightly worse than EasyOCR when the text samples outside the glyph-overlap condition are pooled, suggesting that its overall advantage is largely driven by this author-designed condition. Additional experiments on an external dataset or further analysis would strengthen RTJ’s practical benefits over OCR-based approaches.

Q2. Could the authors clarify the quota-related failures mentioned in the abstract and the caption of Table 3? I understood these as usage limits encountered during the experiments. Please feel free to correct me if I have misunderstood their cause or relevance to moderation performance.

Reviewer_Dx1vOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates a safety moderation gap between the source-code representation of visual artifacts generated by software agents and their rendered appearance. The authors introduce RenderGapBench, consisting of 770 constructed stimuli derived primarily from AdvBench and HarmBench, covering multiple code carriers and obfuscation conditions. The experiments show that several locally deployable safety guards struggle to detect harmful content embedded in obfuscated source code, despite successfully recognizing the corresponding raw text. To address this issue, the authors propose Rasterize–Transcribe–Judge (RTJ), which renders code into images, transcribes visible text using a vision-language model, and applies a text safety classifier.

Critical strengths
  1. Meaningful research problem and controlled evaluation. The paper identifies an important distinction between a safety model's inherent detection capability and the representation it actually receives during deployment. The comparison among code, pixel, and raw-text views provides useful evidence for distinguishing representation-related failures from failures of harmful-content recognition.
  2. Broad empirical investigation and transparent reporting. The authors evaluate multiple model families, obfuscation transformations, and defense strategies. They also report several negative findings and preregistration deviations, including the limitations of canonicalization and the failure of RTJ on pure-geometry symbols. These results provide useful insights into the strengths and limitations of different moderation approaches.
Critical weaknesses
  1. Insufficient validation of semantic preservation in adversarial conditions. A central assumption of the paper is that the applied transformations preserve harmful content that remains recognizable to human observers. However, Appendix K explicitly states that the preregistered human-legibility evaluation was not completed before submission. Furthermore, Appendix B reports that 18 of the 20 primary-source goals contain at least one clipped line after per-glyph rendering, meaning that the displayed content is not always identical to the original text. Although clipping may make the measured code-versus-pixel detection gap conservative, it still raises concerns about the semantic equivalence of the evaluated representations. This is particularly important when interpreting the robustness results under adversarial layouts. The authors should complete the preregistered human-legibility evaluation and report the principal results on unclipped, independently verified, human-readable stimuli.
  2. Insufficient statistical evidence for RTJ's advantage over strong baselines. The paper emphasizes that RTJ maintains at least 60% detection across all evaluated text conditions. However, several important robustness conditions contain only 20 harmful examples, making these estimates statistically uncertain. For example, Table 4 reports 80% detection for RTJ versus 55% for PP-OCRv6 on the glyph-overlap condition, corresponding to 16 versus 11 successful detections out of 20. While this difference is promising, the sample size is insufficient to establish a statistically significant advantage using a two-sided exact McNemar test. Moreover, RTJ does not consistently outperform strong OCR-based alternatives: EasyOCR achieves 65% versus RTJ's 60% under the combined perturbation and 77% versus 73% on naturalistic artifacts. The results support RTJ as a useful approach for certain challenging text layouts, but do not yet establish its robust superiority over existing alternatives. Stronger paired statistical comparisons, confidence intervals, and evaluations at matched false-positive rates would substantially strengthen the contribution.
  3. Limitations in benchmark validity and generalizability. RenderGapBench is primarily composed of programmatically constructed stimuli derived from a limited set of harmful textual goals. Although the benchmark contains 770 stimuli, many share the same underlying semantic content, and some carriers produce identical rendered images. Consequently, the number of constructed stimuli should not be interpreted as the number of independent harmful-content examples. Ground-truth reliability also warrants closer examination. Appendix H reports a 46.9% disagreement rate between the two judges on the second source. While the authors provide human arbitration and sensitivity analyses for part of the evaluation, the independent validation of the more heterogeneous second-source labels remains insufficiently clear. Finally, the evaluation is predominantly text-centric. RTJ achieves 0% detection on the evaluated harmful geometric symbols, and the naturalistic artifacts are produced under controlled attack prompts rather than sampled from deployed agent systems. These limitations do not invalidate the observed rendering gap, but they restrict the extent to which the results establish a generally applicable visual-content moderation solution.
Questions for authors
  1. Semantic preservation and human readability. Can the authors complete the preregistered human-legibility evaluation and report whether the main findings persist on unclipped, human-readable stimuli?
  2. Statistical significance of RTJ improvements. Can the authors provide paired statistical tests and confidence intervals comparing RTJ against the strongest OCR-based and direct-vision baselines under matched false-positive rates?
  3. Benchmark reliability and generalization. How were the substantial judge disagreements on the second source resolved? Can the authors clarify the number of independent semantic examples and provide stronger evidence that the findings generalize beyond constructed text-centric artifacts?

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

Related papers

Discussion (0)

Sign in to comment.