← 研踪 PaperTrailarXiv:2611.24231 · v1
What Survives the Shift from Text to Pixels? Conditional Protection against Prompt Injection in Image-Rendered Documents
首次提交 2026-11-18 19:08
版本更新 2026-11-18 19:08
雷达观测 2026-11-19 15:04
摘要
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.
观测到的版本
相邻记录可能跳过中间版本;首次发现 v2 及以上也会标为版本更新。
v1 · 2026-11-18 19:08 · 最新观测版本
What Survives the Shift from Text to Pixels? Conditional Protection against Prompt Injection in Image-Rendered Documents
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.
查看此版本解读 →查看此版本原文 ↗