acceptodds
Under review as a conference paper at ICLR 2027

The Refraction Gap: How Optical Context Compression Inadvertently Bypasses Safety Alignment

Abstract

To operate within finite context windows, interactive agents increasingly rely on optical context compression by rendering extensive textual inputs and interaction histories into compact visual pages. Existing evaluations of these transformations focus predominantly on capability preservation, examining whether models comprehend and act upon compressed content, while overlooking whether safety alignment is similarly maintained. We show that safety behaviors undergo a severe collapse across this representation boundary. Across five multimodal Readers and representative context compression pipelines, we reveal the Refraction Gap: harmful requests consistently refused in plain text are readily fulfilled when identical content is delivered in optical form. Targeted evaluations demonstrate that this vulnerability cannot be attributed to perceptual degradation: optical contexts preserve semantic fidelity and downstream task utility within 5.1 percentage points of the raw-text baseline, yet trigger unsafe compliance in an average of 47.5% of cases under payload rendering (reaching 73.1%) and 42.8% under memory compaction, compared to merely 4.5% in plain text. Finally, we show that this gap is mitigable prior to generation, where a lightweight 32M-parameter router identifies at-risk payloads before inference and routes them back to text. Taken together, our findings show that optical context compression is not merely an efficiency mechanism; it is a safety-relevant representation transformation that existing evaluation protocols fail to capture.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Optical Context Compression, Agent Security, Multimodal Safety, Safety Alignment, Jailbreak

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_AczSOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies a safety issue introduced by optical context compression in long-context multimodal agents. Existing context compression methods mainly evaluate whether models can preserve semantic understanding and downstream task performance after converting text into visual representations, but they rarely examine whether safety alignment is preserved during this representation shift. The authors identify the Refraction Gap, where identical safety-sensitive content that is refused in plain text can be fulfilled when delivered through optical representations. The paper proposes the Refraction evaluation framework, which compares raw text, optical payload rendering, and visual memory compression under controlled paired settings. Experiments on five multimodal models show that optical compression largely preserves semantic and utility performance while substantially increasing unsafe responses. The authors further introduce GapBridge, a lightweight text-based risk router that detects vulnerable inputs before optical conversion and selectively keeps them in textual form.

Critical strengths
  1. The paper identifies a practical safety failure mode in multimodal agents. Unlike conventional jailbreaks requiring adversarial images, the Refraction Gap emerges from the agent’s own text-to-visual transformation process. This highlights an important limitation of current safety evaluations for optical context compression.

  2. The paired evaluation framework provides strong evidence that capability preservation does not guarantee safety preservation. By controlling content and models while varying only the representation modality, the authors show that optical compression maintains semantic and task performance but significantly weakens safety alignment.

  3. The proposed mitigation demonstrates practical applicability. GapBridge introduces a lightweight pre-conversion risk router that avoids costly model retraining and reduces exposure to the identified vulnerability by selectively keeping risky inputs in text form.

Critical weaknesses
  1. The paper does not sufficiently establish the generality of the Refraction Gap beyond the specific evaluated compression pipelines. The main limitation is that the claimed phenomenon is validated only on two optical transformations (Glyph-based payload rendering and MemOCR-based memory compression) and five open-weight multimodal models. While the results are consistent within this setting, it remains unclear whether the Refraction Gap is a general property of representation changes or a consequence of specific rendering formats, model architectures, or safety training procedures. Broader validation across more diverse compression mechanisms would be necessary to support stronger claims.

  2. The mechanism behind safety degradation remains largely speculative. The paper argues that visual representations fail to activate safety-related internal representations as effectively as text, but the experiments mainly demonstrate behavioral differences rather than explaining their cause. The discussion suggests possible post-training asymmetry between text safety tuning and visual understanding, but no mechanistic analysis, representation probing, or causal intervention is provided. Therefore, the work identifies an important phenomenon but provides limited insight into why the failure occurs.

  3. The proposed GapBridge mitigation is not sufficiently evaluated as a complete defense. GapBridge only detects around two-thirds of vulnerable cases, meaning a substantial fraction of Refraction Gap instances remain undetected. Moreover, the detector is trained using a specific proxy formulation involving harmful intent and structural patterns, which may not capture diverse real-world failure modes. The paper demonstrates feasibility but does not yet establish that the proposed mitigation provides robust protection in realistic agent environments.

Questions for authors
  1. How much of the observed Refraction Gap is caused by the visual representation itself versus specific rendering styles used by Glyph and MemOCR? The paper attributes the safety degradation to the modality transition from text to optical representations, but both evaluated operators introduce specific layout and formatting choices. Could the authors evaluate whether the same vulnerability persists across alternative rendering strategies with different layouts, resolutions, or visual encodings?

  2. Can the authors provide more evidence about the internal cause of the safety alignment failure? The paper suggests that visual pathways may not effectively trigger refusal-related representations, but the current evidence is primarily behavioral. Have the authors investigated whether hidden representations, attention patterns, or safety-related activation features differ between raw text and optical conditions?

  3. How does the Refraction Gap behave in more realistic long-horizon agent interactions? The current evaluation uses isolated prohibited requests under controlled paired settings. In practical agents, harmful information may emerge through multiple interaction turns and accumulated memory. Can the authors evaluate whether the same failure occurs when safety-sensitive information is gradually introduced and later accessed through compressed visual memory?

Reviewer_bc3LOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper identifies the Refraction Gap, a safety alignment failure in optical context compression. The authors show that rendering text into visual representations can substantially increase unsafe compliance while largely preserving semantic understanding and task utility. Unlike conventional multimodal jailbreaks, this failure can arise from an agent's internal compression pipeline without adversarially crafted inputs.

Experiments across five multimodal Readers show that the average attack success rate (ASR) increases from 4.5% under raw text to 47.5% with Glyph and 42.8% with MemOCR. The authors also propose GapBridge, a lightweight 32M-parameter router that identifies approximately two-thirds of vulnerable cases before optical conversion.

Critical strengths
  1. Important system-level safety problem. The paper highlights that semantic and utility preservation do not guarantee alignment preservation. This motivates evaluating safety throughout the context compression pipeline rather than focusing solely on model capabilities.

  2. Consistent empirical observations. Experiments across five Readers, two optical operators, and two safety evaluators show substantial increases in unsafe compliance, providing evidence that the observed vulnerability warrants further investigation.

Critical weaknesses
  1. Unverified paired analysis and overstated causal claims. Appendix C explicitly states that the paired transition counts and associated statistics were reconstructed from aggregate ASRs under assumptions about reverse transitions. Nevertheless, Section 4.2 uses these results to claim a deterministic causal breakdown and rule out sampling artifacts. Aggregate ASRs cannot uniquely determine the paired transition counts or McNemar statistics.

  2. Inconsistency between the threat model and safety evaluation. Section 3 explicitly requires the actionable instruction QQ to remain in plain text, with only the reference payload PP eligible for optical transformation. However, Table 1 describes the alignment evaluation as “Prohibited queries → glyph / memocr,” and Appendix A states that all 520 AdvBench requests are evaluated under three representation conditions. This raises a fundamental ambiguity: are these prohibited requests treated as QQ or PP? If the requests themselves are rendered as images, the evaluation violates the stated constraint on QQ and may primarily demonstrate conventional visual jailbreaks rather than safety failures induced by internal context compression. This ambiguity undermines the validity of the paper's central experimental setup.

Questions for authors
  1. Paired analysis: Why were the transition counts reconstructed rather than computed from item-level logs? Can the authors provide verified transition matrices and recompute the statistical tests?

  2. Experimental protocol: How are prohibited requests represented under raw, Glyph, and MemOCR conditions? Specifically, how is the (Q/P) separation enforced, and does MemOCR involve actual history compaction?

  3. Comprehension control: Can the authors report unsafe transitions restricted to instances where the Reader correctly understands the request in both representations and explicitly refuses it under raw text?

  4. Mitigation effectiveness: What are the end-to-end ASR and false-positive rate after applying GapBridge? How does it compare with a standard text-based harmful-content classifier?

Reviewer_ns6KOpenReview ↗
Rating
3(+1)
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies safety degradation under optical context compression. Across five multimodal models, it reports substantially higher harmful response rates when inputs are rendered through Glyph or MemOCR rather than provided as plain text. It also proposes GapBridge, a lightweight classifier that routes potentially risky inputs back to text.

Critical strengths
  1. The paper raises a relevant concern: preserving comprehension and task performance after context compression does not necessarily preserve safety.
  2. The reported increase in harmful responses appears across multiple models, two rendering methods, and two automated evaluators, making the observation worth investigating.
Critical weaknesses
  1. The agent scenario described in the paper is not actually tested. The introduction describes a multi-turn conversation where automatic history compaction causes the model to switch from refusal to an unsafe response. The experiments, however, compare standalone harmful requests in text and image form, without testing the full conversation and compaction process. This is a reasonable motivation, but presenting it as an established finding overstates the evidence.

  2. The difference from existing typographic jailbreak work remains unclear. Rendering harmful text as images to elicit unsafe responses is not a new finding. This paper emphasizes that the images are generated internally by an agent, but does not test the agent’s actual operation or compare its rendering methods with ordinary text screenshots. It is therefore unclear what new insight comes specifically from context compression.

  3. Some conclusions lack direct evidence. Similar performance on benign QA tasks does not show that the model accurately understood each harmful input, so changes in perception or understanding cannot simply be ruled out. GapBridge also reports only detection recall, without showing false-positive rates, harmful response rates after routing, or the remaining token savings. These results are insufficient to judge whether the defense is practical.

Questions for authors

see weakness above

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.