Rendering untrusted documents as page images cuts append and format attacks on six models, but Qwen3-VL-8B still executes image-borne answer replacement at 77.96%
Synopsis
Rendering untrusted documents as page images is widely used against indirect prompt injection, but only its average effect on attack success was known. This study pairs text and page-image versions of 191 English question-answering sources across six models and measures three injected operations separately: appending a marker, formatting the answer, and replacing it, alongside clean question-answering utility. Rendering lowers append and format execution for every model, yet Qwen3-VL-8B still executes image-borne answer replacement at 77.96%, so the protection is conditional.
Interpretation
After rendering to page images, the rate at which injected commands are executed falls across the six models, but the size of the drop depends on the model and on what the command asks the model to do. Earlier results established that rendering lowers attack execution on average, without separating appending a marker line, formatting the answer, and replacing the answer with a fixed phrase. On 191 English LongBench question-answering sources, image attack success for append and format is lower than the text version for all six models; Gemini 3.1 Flash-Lite still executes roughly 50% of image append and format attacks while the other models stay below 12%, and Qwen3-VL-8B executes image answer replacement at 77.96%.
Clean question-answering utility is largely preserved, but for some models it drops sharply when a question requires combining two evidence spans. Utility ratios on the original questions reflect only single-span evidence and do not characterize the cost of integrating evidence across a rendered page. Utility ratios on original questions range from 93.31% to 101.61% across the six models, while Qwen3-VL-8B falls to 61.99% on the added integration questions, and Haiku 5.5, Gemini 3.1 Flash-Lite and GLM-5.3-Flash show no loss on integration questions.
Failed image attacks cannot all be attributed to the model being unable to read the command, since models often transcribe and identify image-borne commands they do not execute. The earlier explanation leaned on rendering making text hard to recognize, thereby weakening the injection. For Qwen3-VL-8B, among sources with successful image reporting, execution is 0.00% for append and 3.70% for format; Gemini 3.1 Flash-Lite reports and identifies more than 90% of image append and format commands yet executes far fewer; replacing each page with its OCR transcript yields 41.80% execution versus 70.96% for the original text and 1.11% for images.
Inside Qwen3-VL-8B, a text command acts through its own positions, whereas an image command and its target are also spread across visual tokens elsewhere on the instruction page. Earlier work localized late-layer states governing injection compliance in text models, but it was unclear which regions encode the semantics of a rendered command. Patching the attack region induces all three operations, while random and position-shifted controls elicit none in the scored responses; blocking the attack region eliminates text execution but barely affects images, and only blocking the whole page eliminates image execution, with modality-by-blocking interactions of +84.38 points for append and +80.21 for format.
Perspective
The findings rest on paired text-image comparisons over 191 English LongBench question-answering sources (80 Qasper, 69 MultiFieldQA-en, 42 HotpotQA) and six models; presentation experiments use 90 of those sources, and the mechanistic analysis is confined to the open-weight Qwen3-VL-8B. The authors accordingly propose that rendering-based defenses report execution by model, operation and attack budget, evaluated together with utility on questions that integrate evidence.
The authors note that low reporting or identification rates may reflect failure to follow the diagnostic task or to reproduce the exact target rather than unread content; recognition errors mainly reduce exact reproduction while their effect on operation execution remains unresolved; image append intervention effects are smaller and less precisely resolved; and in images the bound target code is decodable from attack-region states, yet those sites do not determine which code the response produces.
