One Perturbation, Two Failure Modes: Probing VLM Safety via Embedding-Guided Typographic Perturbations

• URL: http://arxiv.org/abs/2604.25102v1
• Date: Tue, 28 Apr 2026 01:21:47 GMT
• Title: One Perturbation, Two Failure Modes: Probing VLM Safety via Embedding-Guided Typographic Perturbations
• Abstract summary: Typographic prompt injection exploits vision language models' (VLMs) ability to read text rendered in images.<n>We show how to maximize image text embedding similarity under bounded $ell_infty$ perturbations.<n>Experiments on GPT-4o, Claude Sonnet 4.5, Mistral-Large-3, and Qwen3-VL confirm that optimization recovers readability and reduces safety aligned refusals.
• Score: 0.8116687935773981
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Typographic prompt injection exploits vision language models' (VLMs) ability to read text rendered in images, posing a growing threat as VLMs power autonomous agents. Prior work typically focus on maximizing attack success rate (ASR) but does not explain \emph{why} certain renderings bypass safety alignment. We make two contributions. First, an empirical study across four VLMs including GPT-4o and Claude, twelve font sizes, and ten transformations reveals that multimodal embedding distance strongly predicts ASR ($r{=}{-}0.71$ to ${-}0.93$, $p{<}0.01$), providing an interpretable, model agnostic proxy. Since embedding distance predicts ASR, reducing it should improve attack success, but the relationship is mediated by two factors: perceptual readability (whether the VLM can parse the text) and safety alignment (whether it refuses to comply). Second, we use this as a red teaming tool: we directly maximize image text embedding similarity under bounded $\ell_\infty$ perturbations via CWA-SSA across four surrogate embedding models, stress testing both factors without access to the target model. Experiments across five degradation settings on GPT-4o, Claude Sonnet 4.5, Mistral-Large-3, and Qwen3-VL confirm that optimization recovers readability and reduces safety aligned refusals as two co-occurring effects, with the dominant mechanism depending on the model's safety filter strength and the degree of visual degradation.

Related papers

• Quality-Diversity Evolution for Discovering Diverse Vulnerabilities in LLM Safety [0.0]
  We introduce a quality-diversity evolutionary framework that operates at the semantic level.<n>We maintain a diverse archive of attacks across behavioral dimensions.<n>The semantic representation produces interpretable attacks that reveal systematic, model-specific weaknesses.
  arXiv  Detail & Related papers  (2026-05-30T16:40:24Z)
• Persona-Model Collapse in Emergent Misalignment [0.0]
  Fine-tuning large language models on narrow data with harmful content produces broadly misaligned behavior on unrelated prompts.<n>We test this hypothesis using two metrics: moral susceptibility (S) and moral robustness (R)<n>These metrics formalize the model's ability to differentiate characters (S) and its consistency when simulating a given one (R)
  arXiv  Detail & Related papers  (2026-05-13T00:48:57Z)
• Reading Between the Pixels: Linking Text-Image Embedding Alignment to Typographic Attack Success on Vision-Language Models [4.577407934990345]
  We study typographic prompt injection attacks on vision-language models (VLMs), where adversarial text is rendered as images to bypass safety mechanisms.<n>In practice, the attack surface is heterogeneous: adversarial text appears at varying font sizes and under diverse visual conditions.
  arXiv  Detail & Related papers  (2026-04-14T06:59:27Z)
• Multi-Faceted Attack: Exposing Cross-Model Vulnerabilities in Defense-Equipped Vision-Language Models [54.61181161508336]
  We introduce Multi-Faceted Attack (MFA), a framework that exposes general safety vulnerabilities in leading defense-equipped Vision-Language Models (VLMs)<n>The core component of MFA is the Attention-Transfer Attack (ATA), which hides harmful instructions inside a meta task with competing objectives.<n>MFA achieves a 58.5% success rate and consistently outperforms existing methods.
  arXiv  Detail & Related papers  (2025-11-20T07:12:54Z)
• VERA-V: Variational Inference Framework for Jailbreaking Vision-Language Models [19.867040067010674]
  We introduce VERA-V, a variational inference framework that recasts multimodal jailbreak discovery as learning a joint posterior distribution over paired text-image prompts.<n>We train a lightweight attacker to approximate the posterior, allowing efficient sampling of diverse jailbreaks.<n>Experiments on HarmBench and HADES benchmarks show that VERA-V consistently outperforms state-of-the-art baselines.
  arXiv  Detail & Related papers  (2025-10-20T17:12:10Z)
• SaFeR-VLM: Toward Safety-aware Fine-grained Reasoning in Multimodal Models [66.71948519280669]
  Multimodal Large Reasoning Models (MLRMs) demonstrate impressive crossmodal reasoning but often amplify safety risks under adversarial prompts.<n> Existing defenses mainly act at the output level and do not constrain the reasoning process, leaving models to implicit risks.<n>We propose SaFeR-VLM, which integrates four components and supports dynamic and interpretable safety decisions beyond surface-level filtering.
  arXiv  Detail & Related papers  (2025-10-08T10:39:12Z)
• JPS: Jailbreak Multimodal Large Language Models with Collaborative Visual Perturbation and Textual Steering [73.962469626788]
  Jailbreak attacks against multimodal large language Models (MLLMs) are a significant research focus.<n>We propose JPS, underlineJailbreak MLLMs with collaborative visual underlinePerturbation and textual underlineSteering.
  arXiv  Detail & Related papers  (2025-08-07T07:14:01Z)
• Retention Score: Quantifying Jailbreak Risks for Vision Language Models [60.48306899271866]
  Vision-Language Models (VLMs) are integrated with Large Language Models (LLMs) to enhance multi-modal machine learning capabilities.<n>This paper aims to assess the resilience of VLMs against jailbreak attacks that can compromise model safety compliance and result in harmful outputs.<n>To evaluate a VLM's ability to maintain its robustness against adversarial input perturbations, we propose a novel metric called the textbfRetention Score.
  arXiv  Detail & Related papers  (2024-12-23T13:05:51Z)
• Dissecting Adversarial Robustness of Multimodal LM Agents [70.2077308846307]
  We manually create 200 targeted adversarial tasks and evaluation scripts in a realistic threat model on top of VisualWebArena.<n>We find that we can successfully break latest agents that use black-box frontier LMs, including those that perform reflection and tree search.<n>We also use ARE to rigorously evaluate how the robustness changes as new components are added.
  arXiv  Detail & Related papers  (2024-06-18T17:32:48Z)
• Emulated Disalignment: Safety Alignment for Large Language Models May Backfire! [65.06450319194454]
  Large language models (LLMs) undergo safety alignment to ensure safe conversations with humans. This paper introduces a training-free attack method capable of reversing safety alignment. We name this method emulated disalignment (ED) because sampling from this contrastive distribution provably emulates the result of fine-tuning to minimize a safety reward.
  arXiv  Detail & Related papers  (2024-02-19T18:16:51Z)
• VQAttack: Transferable Adversarial Attacks on Visual Question Answering via Pre-trained Models [58.21452697997078]
  We propose a novel VQAttack model, which can generate both image and text perturbations with the designed modules. Experimental results on two VQA datasets with five validated models demonstrate the effectiveness of the proposed VQAttack.
  arXiv  Detail & Related papers  (2024-02-16T21:17:42Z)
• FigStep: Jailbreaking Large Vision-Language Models via Typographic Visual Prompts [14.33139608409507]
  We propose FigStep, a straightforward yet effective black-box jailbreak algorithm against LVLMs.<n>FigStep converts the prohibited content into images through typography to bypass the safety alignment.<n>Our work reveals that current LVLMs are vulnerable to jailbreak attacks, which highlights the necessity of novel cross-modality safety alignment techniques.
  arXiv  Detail & Related papers  (2023-11-09T18:59:11Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.