Back to AI News
arxiv

What Survives the Shift from Text to Pixels? Conditional Protection against Prompt Injection in Image-Rendered Documents

This arXiv paper studies indirect prompt injection when multimodal language models answer questions over external documents, comparing text and page-image versions of 191 English QA sources across six models. Rendering untrusted content as images reduces marker-appending and answer-formatting execution for every model, but Qwen3-VL-8B still executes image-borne answer replacement at 77.96%. Protection varies by target type, rendering settings, wording, and repetition, and an eight-candidate attack search largely closes the image deficit for Qwen and Gemini but not GPT-5-nano. Clean-question utility ranges from 93.31% to 101.61%, though Qwen drops to 61.99% when answers require combining two evidence spans.

AI Brief

It provides concrete measurements of when image rendering mitigates prompt injection and when it fails, helping teams evaluate multimodal document security by model, operation, and attack budget.

arXiv:2611.24231v1 Announce Type: cross Abstract: Multimodal language models increasingly answer questions over external documents, exposing them to indirect prompt injection: commands planted in untrusted content to hijack the response. Rendering untrusted content as page images lowers attack success on average, leaving open which injected operations survive rendering, at what utility cost, and whether failed image attacks reflect unread commands or commands read but not followed. We pair text and page-image versions of 191 English question-answering sources across six models, measuring marker appending, answer formatting, and answer replacement alongside clean Question-Answering utility. Rendering reduces append and format execution for every model, yet Qwen3-VL-8B still executes image-borne replacement at 77.96\%. Protection also depends on target type, rendering settings, wording, and repetition, and 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. Utility ratios on original questions range from 93.31\% to 101.61\%, but Qwen falls to 61.99\% when answers combine two evidence spans. Illegibility cannot explain all failed attacks: models often transcribe and identify image-borne commands they do not execute, and trusted user authorization raises execution on unchanged pages. Inside Qwen, patching early attack-region states induces all three operations, more weakly in images. Attention blocking suggests why: a text command acts through its own positions, whereas visual encoding spreads an image command and its target across the instruction page. Rendering thus offers conditional protection, motivating evaluations that report security by model, operation, and attack budget alongside utility.
Be the first to rate this article.

Comments & Questions (0)

Captcha
Please be respectful — let's keep the conversation friendly.

No comments yet

Be the first to comment!