Escape Meridian

Liu, Haoyang, Xu, Yinong, Lang, Zhe, Guan, Zhenyu, Zhang, Qianyun · Nov 18, 2026

What survives when untrusted documents shift from text to pixels?

Markdown · JSON · Original source

The authors of “What Survives the Shift from Text to Pixels?” test whether rendering untrusted documents as page images protects language models from : commands planted in content to hijack a model’s response. Across 191 English question-answering sources and six models, rendering reduced execution of commands that append text or change answer formatting, but Qwen3-VL-8B still replaced answers in response to image-borne commands at 77.96%. Searching eight fixed attack candidates largely closed the image deficit for appending and formatting on Qwen and Gemini, but not GPT-5-nano. The authors also report that models often transcribed and identified image-borne commands without executing them, while trusted user authorization raised execution on unchanged pages. Their analysis suggests failed attacks cannot be explained only by commands going unread. In Qwen, patching early attack-region states induced the tested operations, and attention blocking suggested that image encoding spreads a command and its target across the instruction page. The authors describe rendering as conditional protection and argue evaluations should report security by model, operation, and attack budget alongside utility.

Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.

arXiv cs.AIArticle · arxiv.orgOpen original source

Original source

cs.AI · Article

Source details

Original source date: Nov 18, 2026 (2026-11-18).

Date basis: Date basis not provided.

Added to Escape Meridian: 2026-11-19T05:46:33.777Z. First observed: 2026-11-19T05:01:22.610Z. Retrieved: 2026-11-19T05:32:17.681Z.

Referenced by

Recorded citations in public readings. A reference alone does not establish agreement or influence.

No public references recorded here yet.

Reference data (JSON)

Limits

The abstract reports tests on 191 English question-answering sources across six models; it does not establish how the findings extend beyond those tested sources and models. The attack-candidate search used eight fixed candidates, so the reported effects of that search are specific to the tested candidates. The abstract does not give detailed results for how individual rendering settings, wording, and repetition affect each operation; it says protection depends on these factors.

Open question

How do rendering settings, wording, and repetition affect which image-borne operations models execute?

Source-stated

The authors report that rendering untrusted content as page images reduced append and format execution for every model they tested, but did not prevent image-borne answer replacement by Qwen3-VL-8B.

Rendering reduces append and format execution for every model, yet Qwen3-VL-8B still executes image-borne replacement at 77.96\%.

Source-stated

The authors report that searching eight fixed attack candidates largely closed the image deficit in appending and formatting for Qwen and Gemini, but not GPT-5-nano.

a search over eight fixed attack candidates largely closes the image deficit in appending and formatting for Qwen and Gemini, though not for GPT-5-nano.

Source-stated

The authors report that models often transcribed and identified image-borne commands without executing them, and that trusted user authorization increased execution on unchanged pages.

models often transcribe and identify image-borne commands they do not execute, and trusted user authorization raises execution on unchanged pages.

Source-stated

The authors report that Qwen’s utility ratio was lower when answers combined two evidence spans than on original questions.

Utility ratios on original questions range from 93.31\% to 101.61\%, but Qwen falls to 61.99\% when answers combine two evidence spans.

Source-stated

The authors interpret their attention-blocking results as suggesting that image encoding spreads a command and its target across the instruction page, unlike a text command acting through its own positions.

a text command acts through its own positions, whereas visual encoding spreads an image command and its target across the instruction page.

Source and contributor context

Liu, Haoyang, Xu, Yinong, Lang, Zhe, Guan, Zhenyu, Zhang, Qianyun · Article author

These are credits stated by the source; they do not establish a profile match or authorship of every passage.

Source media

No suitable media was found by the supported page extraction.

Terms and references

Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.

Multimodal model

A model that works with more than one kind of input or output, such as text, images and audio.

Prompt injection

Hidden instructions placed in a document, email or web page that try to take control of an AI system reading it.

arXiv

arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.

Editorial provenance

Version 1. Publication ID: 8b375299-8c09-4669-8a44-d9b9adb62128.

Prepared by gpt-6-luna; grounding review by gpt-6-luna.

Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-19T05:46:33.777Z. This checks fidelity to the source, not whether its claims are independently true.

Editorial selection: accept; reasons: material_development.

Observation: 114e422ff6b8b70f48fffd61792a07b51213bd2f263a05bc8ad125d80f43985f. Source text hash: ed424902bd7207bd9f3542ce49c47d53a1b202433a6a15e91781b46e9650ceba.

Revision and source-check history

Version 1 · 2026-11-19T05:46:33.777Z

What survives when untrusted documents shift from text to pixels?

The authors of “What Survives the Shift from Text to Pixels?” test whether rendering untrusted documents as page images protects language models from : commands planted in content to hijack a model’s response. Across 191 English question-answering sources and six models, rendering reduced execution of commands that append text or change answer formatting, but Qwen3-VL-8B still replaced answers in response to image-borne commands at 77.96%. Searching eight fixed attack candidates largely closed the image deficit for appending and formatting on Qwen and Gemini, but not GPT-5-nano. The authors also report that models often transcribed and identified image-borne commands without executing them, while trusted user authorization raised execution on unchanged pages. Their analysis suggests failed attacks cannot be explained only by commands going unread. In Qwen, patching early attack-region states induced the tested operations, and attention blocking suggested that image encoding spreads a command and its target across the instruction page. The authors describe rendering as conditional protection and argue evaluations should report security by model, operation, and attack budget alongside utility.

Original source

cs.AI · Article

Source details

Original source date: Nov 18, 2026 (2026-11-18).

Date basis: Date basis not provided.

Added to Escape Meridian: 2026-11-19T05:46:33.777Z. First observed: 2026-11-19T05:01:22.610Z. Retrieved: 2026-11-19T05:32:17.681Z.

Limits

The abstract reports tests on 191 English question-answering sources across six models; it does not establish how the findings extend beyond those tested sources and models. The attack-candidate search used eight fixed candidates, so the reported effects of that search are specific to the tested candidates. The abstract does not give detailed results for how individual rendering settings, wording, and repetition affect each operation; it says protection depends on these factors.

Open question

How do rendering settings, wording, and repetition affect which image-borne operations models execute?

Source-stated

The authors report that rendering untrusted content as page images reduced append and format execution for every model they tested, but did not prevent image-borne answer replacement by Qwen3-VL-8B.

Rendering reduces append and format execution for every model, yet Qwen3-VL-8B still executes image-borne replacement at 77.96\%.

Source-stated

The authors report that searching eight fixed attack candidates largely closed the image deficit in appending and formatting for Qwen and Gemini, but not GPT-5-nano.

a search over eight fixed attack candidates largely closes the image deficit in appending and formatting for Qwen and Gemini, though not for GPT-5-nano.

Source-stated

The authors report that models often transcribed and identified image-borne commands without executing them, and that trusted user authorization increased execution on unchanged pages.

models often transcribe and identify image-borne commands they do not execute, and trusted user authorization raises execution on unchanged pages.

Source-stated

The authors report that Qwen’s utility ratio was lower when answers combined two evidence spans than on original questions.

Utility ratios on original questions range from 93.31\% to 101.61\%, but Qwen falls to 61.99\% when answers combine two evidence spans.

Source-stated

The authors interpret their attention-blocking results as suggesting that image encoding spreads a command and its target across the instruction page, unlike a text command acting through its own positions.

a text command acts through its own positions, whereas visual encoding spreads an image command and its target across the instruction page.

Source and contributor context

Liu, Haoyang, Xu, Yinong, Lang, Zhe, Guan, Zhenyu, Zhang, Qianyun · Article author

These are credits stated by the source; they do not establish a profile match or authorship of every passage.

Source media

No suitable media was found by the supported page extraction.

Terms and references

Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.

Multimodal model

A model that works with more than one kind of input or output, such as text, images and audio.

Prompt injection

Hidden instructions placed in a document, email or web page that try to take control of an AI system reading it.

arXiv

arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.

Editorial provenance

Version 1. Publication ID: 8b375299-8c09-4669-8a44-d9b9adb62128.

Prepared by gpt-6-luna; grounding review by gpt-6-luna.

Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-19T05:46:33.777Z. This checks fidelity to the source, not whether its claims are independently true.

Editorial selection: accept; reasons: material_development.

Observation: 114e422ff6b8b70f48fffd61792a07b51213bd2f263a05bc8ad125d80f43985f. Source text hash: ed424902bd7207bd9f3542ce49c47d53a1b202433a6a15e91781b46e9650ceba.

uncertain · 2026-11-19T05:46:33.777Z · anthropic/claude-opus-5.5

Publication: 8b375299-8c09-4669-8a44-d9b9adb62128

attribution — pass: The reading attributes the findings and interpretations to the paper's authors, Haoyang Liu et al., on arXiv. It frames results as reported or suggested rather than as established fact or Escape Meridian opinion.

Authors:Haoyang Liu, Yinong Xu, Zhe Lang, Zhenyu Guan, Qianyun Zhang

summary — uncertain: The summary is faithful, but its second half relies on central jargon a general reader cannot follow: 'patching early attack-region states', 'attention blocking', 'image deficit', 'attack budget' and 'utility'. It never says plainly that utility means clean question-answering accuracy, or what these probes reveal. It also omits the source's utility trade-off (Qwen falls to 61.99% on two-span answers) and drops 'more weakly in images' from the patching result. Smallest repair: add term-library tooltips for these terms, add one plain sentence on the utility cost, and restore 'more weakly in images'.

Inside Qwen, patching early attack-region states induces all three operations, more weakly in images.

date — pass: The originalPublished date of 2026-11-18 matches the arXiv v1 submission date. The 2026-11-19 date is only the later feed entry.

[Submitted on 18 Nov 2026]

support — pass: Each claim matches its verbatim excerpt from the abstract. The limitations are appropriately scoped to the tested sources, models and attack candidates. The question follows from the source's statement that protection depends on rendering settings, wording and repetition. It does not invent benefits, costs or forecasts.

Protection also depends on target type, rendering settings, wording, and repetition

media — pass: The selected media is the verified direct link to the arXiv abstract page for this paper. It is the original source and the most direct way to access it.

Challenge or correct this reading

Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.