AI & ML
What Survives the Shift from Text to Pixels? Conditional Protection against Prompt Injection in Image-Rendered Documents
arXiv cs.CR0 views
arXiv:2611.24231v1 Announce Type: new 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.
Source:https://arxiv.org/abs/2611.24231PreviousErased but Recoverable: On the Fragility of Concept Erasure Under Fine-TuningNextPMKPIR: Parallel Multi-Keyword PIR with Interleaved Batching for Private Boolean Retrieval
Related posts
CaMBRAIN: Real-time, Continuous EEG Inference with Causal State Space Models
0arXiv cs.HC
How Far Does Online RL Generalize in Mobile GUI Agents? A Controlled Study with AndroidWorld-Generalization
0arXiv cs.HC
From Consensus-Based Optimization to Evolution Strategies: Proof of Global Convergence
0arXiv cs.NA
Unconditionally $L^2$-stable explicit--implicit--null filtered Runge--Kutta methods for general nonlinear degenerate parabolic equations
0arXiv cs.NA
CS-PIKAN: Curvilinear Separable Physics-Informed Kolmogorov-Arnold Networks with Autograd-Based Metric Evaluation and Certified Error Bounds for PDEs on Complex Domains
0arXiv cs.NA
Fourier--Galerkin Methods for Subwavelength Resonances in two-dimensional Acoustic Metamaterials
0arXiv cs.NA