Skip to content
InfoResearchPreprintLLM-specific

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

Published
Record updated
View JSON

Summary

This study compares text and page-image versions of 191 English question-answering sources across six multimodal models to test indirect prompt injection. Rendering untrusted content as images reduces marker appending and answer formatting 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 Qwen falls to 61.99% utility when answers combine two evidence spans.