cs.AINov 12, 2026

Scale-Induced Semantic Shift in Visual Prompt Injection Against VLMs and Web Agents

Authors: Yiheng Zhao, Runze Ji, Chenyu Yi, Yi Yu, Bingquan Shen, Alex Kot, Xudong Jiang

Organizations: Nanyang Technological University, Singapore · DSO National Laboratories, Singapore

Abstract

Visual content in multimedia systems may be resized, compressed, or re-rendered before it is interpreted by a vision-language model. We study a security failure mode in which these transformations change the machine-readable semantics of an image: a textual instruction is not recovered at a 512×512512\times512 presentation resolution under our evaluation protocol but becomes recoverable after aggressive downsampling. We formulate this behavior as Scale-Induced Semantic Shift (SISS) and develop a closed-loop black-box framework that combines dual-scale verification, selective self-hardening, shallow-layer low-rank adaptation, and pixel-space noise regularization. On a fixed 300-case benchmark, the third SISS refinement round obtains a 58.75%58.75\% average attack success rate across four closed-source vision-language models, compared with 41.08%41.08\% for the strongest generation baseline. A controlled sweep bounds the effect to aggressive compression: average success decreases to 17.68%17.68\%, 6.85%6.85\%, and 5.60%5.60\% at 128×128128\times128, 224×224224\times224, and 336×336336\times336, respectively. In a downstream web-agent study, the generated images increase hidden-instruction perception, but absolute behavioral compromise remains about 4%4\%. Generic input transformations provide little protection in a separate defense protocol, whereas the proposed ScaleGuard consistency detector reduces non-adaptive success from 61.4%61.4\% to 24.3%24.3\% and one-round defense-aware success from 42.5%42.5\% to 15.0%15.0\%, with no observed rejection among matched clean controls. The results characterize a bounded but reproducible semantic inconsistency in multimedia preprocessing and motivate scale-consistent visual-language inference.

Figures & tables

Appendix figures & tables14 assets

Supplementary material from the paper’s appendix.

Appendix

Explore similar work

Sep 29, 2026

It Takes Little to Rewrite Perception: Targeted Semantic Substitution in Vision-Language Models at ε≤4/255ε\leq 4/255

Vision Language Models (VLMs) are widely deployed in safety-critical scenarios, and understanding to which extent they can be controlled by adversarial perturbation is a prerequisite for evaluating their trustworthiness. Existing representation-alignment attacks, which make a VLM perceive a target image, achieve limited success at ε≤4/255\varepsilon \leq 4/255. Therefore, VLMs seems robust to perturbations in this range. We show that this robustness does not hold, as targeted semantic substitution succeeds within the same range. Specifically, we align each stream of the source image with its counterpart in the target image in the victim VLM's post-merger token space, operating under a white-box threat model. We evaluate under a strict success criterion, requiring the model to simultaneously name the target, confirm its presence, and deny the source. In images, target semantics appear at ε=2/255\varepsilon = 2/255 and complete replacement reaches 38% at ε=4/255\varepsilon = 4/255. On video, complete replacement reaches 35.9% at ε=1/255\varepsilon = 1/255. We also observe a phenomenon of semantic fusion, where Large Language Model (LLM) rationalizes contradictory visual signals into a coherent narrative.
Nov 17, 2026

Reading Scene States: Uncovering Safety Failures in Vision-Language Models

Vision-language models (VLMs) can provide harmful assistance by interpreting visual events, even without an explicit request for the underlying harmful objective. We identify this scene-state safety vulnerability and introduce SceneSafetyBench, comprising 375 image--question pairs spanning Initial, Medium, and Ultimate scene states and five visual presentations. Across eight VLMs, scene-state inputs elicit harmful assistance with attack success rates (ASRs) of 45.6--72.3%. Comparative analysis across three VLMs shows that scene-state inputs most closely resemble QR-related images in their layer-wise safety-related signal profiles, while exhibiting lower average late-layer readouts and distinct signal dynamics during generation. To turn this vulnerability into a systematic attack, we propose S3TA, which translates harmful objectives into scene states, queries image--question pairs independently, and assembles the responses. On seven MM-SafetyBench categories across eight open-weight and commercial VLMs, S3TA achieves a model-average ASR of 71.2%, exceeding the strongest baseline by 6.1 percentage points. Text-side filters and visual moderation provide limited protection. These findings highlight the need to safeguard the assistance models infer through scene-grounded reasoning, beyond screening input content.
Sep 2, 2026

Beyond the Verdict: Evidence-Aligned Evaluation of Visual Prompt-Injection Guardrails

Verdict-only evaluation does not reveal whether a vision-language model (VLM) used the visual evidence that should support its decision. We study this problem in web-agent guardrails, where a VLM judges whether on-screen text conflicts with a user instruction. We introduce Mind2Web-Injection, a benchmark of 9,954 instruction-screenshot pairs with instruction-relative labels, pixel-exact evidence boxes, and matched image-side counterfactuals. Across six VLMs, two models with nearly identical average precision differ ninefold in Evidence-Aligned Detection (EAD), the fraction of attacks both detected and correctly localized. To test whether a verdict depends on the command cited as evidence, we replace the instruction with one that endorses that command. Qwen3-VL-32B, the strongest open-weight localizer, returns aligned in only 58.7% of cases, whereas GPT-5.6-luna does so in 99.9%. To diagnose these failures, we propose two training-free interventions. ReadGate improves grounding without changing verdicts, while CmdCompare tests whether explicit instruction-command comparison resolves instruction-side inconsistency. These results motivate reporting verdict correctness, evidence localization, and counterfactual responsiveness separately.