Qwen3-VL-8B image-rendered prompt injection rises 25+ pp with higher resolution or larger font
Auto-published from arXiv:2611.24231 by the mine-arxiv pipeline. Reviewed by an LLM judge (Sonnet) against the archive bar — see CONTRIBUTING. Notes: cleared review (confidence 0.82, flags: [no-prompt-excerpt])
Category
multimodal
Model
Qwen3-VL-8B-Instruct
Surface
VLM input — document QA with page images at varied rendering settings
Setup
The same fixed-phrase replacement attack is embedded in LongBench sources rendered to PNG across a 3×3 grid: export widths of 704, 876, and 1280 pixels crossed with CSS font sizes of 20, 26, and 32 pt. System prompt, user question, source content, and attack wording are held fixed; only rendering parameters vary. Measured under prompt wrappers W1 and W2. No verbatim prompt excerpt is provided in the body; the rendering grid and scoring procedure are described in Section 5.2.
Reproduction anchor
benchmark: LongBench 90-source subset (30 per dataset); model Qwen3-VL-8B-Instruct; full 3×3 resolution×font grid specified in Section 5.2; Figure 2(a) reports results. No code URL visible in paper body.
Observed behavior
Enlarging font size from 20 to 32 pt at export width 876 px, or widening the page from 704 to 1280 px at font size 26 pt, each raised attack execution by more than 25 percentage points. At the coarsest setting clean QA F1 also dropped (42.21 vs 48.72 at default), creating a utility–security tradeoff where low attack success at coarse settings is purchased at a meaningful answer-quality cost.
Expected behavior
A rendering-based defense should provide consistent protection across plausible rendering configurations; moderate accessibility-motivated increases to font size or resolution should not dramatically increase injection vulnerability.
Reproducibility
medium
Threat model
A defender who deploys image-rendering as a prompt-injection mitigation may inadvertently choose rendering parameters optimized for readability (larger fonts, higher resolution) that largely negate the security benefit. Attackers who know the rendering settings used in a production pipeline can predict and exploit the resulting vulnerability surface.
Novelty
Establishes a quantitative tradeoff surface between rendering legibility (utility) and injection resistance (security), showing defenders cannot freely tune rendering parameters without re-evaluating security at each setting.
Source
- arXiv: 2611.24231
- PDF: https://arxiv.org/pdf/2611.24231v1
- Categories: cs.CR, cs.AI
- Authors: Haoyang Liu, Yinong Xu, Zhe Lang, Zhenyu Guan, Qianyun Zhang
Triage notes (auto)
- paperType:
benchmark - estimatedCaseCount: 3
- triage reason: Systematic empirical study measuring which prompt injection attack types (marker appending, answer formatting, answer replacement) survive image rendering across 6 VLMs with concrete success rates (e.g., Qwen3-VL-8B at 77.96% replacement success). Reproducible failures at model level; not a pure defense paper since it shows rendering incompletely protects.
0 comments
─────────────────────────────────────────────────────────────────────
// no comments yet