arXiv:2511.16203v3 Announce Type: replace 
Abstract: Vision-Language-Action models (VLAs) have recently demonstrated remarkable progress in embodied environments, enabling robots to perceive, reason, and act through unified multimodal understanding. Despite their impressive capabilities, the adversarial robustness of these systems remains largely unexplored, especially under realistic multimodal and black-box conditions. Existing studies mainly focus on single-modality perturbations and overlook the cross-modal misalignment that fundamentally affects embodied reasoning and decision-making. In this paper, we introduce VLA-Fool, a comprehensive study of multimodal adversarial robustness in embodied VLA models under both white-box and black-box settings. VLA-Fool unifies three levels of multimodal adversarial attacks: (1) textual perturbations through gradient-based and prompt-based manipulations, (2) visual perturbations via patch and noise distortions, and (3) cross-modal misalignment attacks that intentionally disrupt the semantic correspondence between perception and instruction. We further incorporate a VLA-aware semantic space into linguistic prompts, developing the first automatically crafted and semantically guided prompting framework. Experiments on the LIBERO benchmark using a fine-tuned OpenVLA model reveal that even minor multimodal perturbations can cause significant behavioral deviations, demonstrating the fragility of embodied multimodal alignment.

تسلط الأبحاث الحديثة الضوء على نقاط الضعف في نماذج اللغة-الرؤية-الإجراء (VLA) أمام الهجمات العدائية متعددة الوسائط، كما هو موضح في الدراسة 'When Alignment Fails: Multimodal Adversarial Attacks on Vision-Language-Action Models.' تقدم الدراسة VLA-Fool، وهو إطار عمل يفحص قوة النماذج VLA أمام الهجمات العدائية في ظروف كل من الصندوق الأبيض والصندوق الأسود، مع التركيز على عدم التوافق بين الوسائط الذي يؤثر على اتخاذ القرار.

Investigaciones recientes destacan las vulnerabilidades de los modelos de Acción-Lenguaje-Visión (VLA) ante ataques adversariales multimodales, como se presenta en el estudio 'When Alignment Fails: Multimodal Adversarial Attacks on Vision-Language-Action Models.' El estudio introduce VLA-Fool, un marco que examina la robustez adversarial de los VLA en condiciones tanto de caja blanca como de caja negra, centrándose en el desajuste intermodal que impacta la toma de decisiones.

Des recherches récentes mettent en lumière les vulnérabilités des modèles Vision-Language-Action (VLA) face aux attaques adversariales multimodales, comme le montre l'étude 'When Alignment Fails: Multimodal Adversarial Attacks on Vision-Language-Action Models.' L'étude présente VLA-Fool, un cadre qui examine la robustesse adversariale des VLA dans des conditions à la fois de boîte blanche et de boîte noire, en se concentrant sur le désalignement intermodal qui impacte la prise de décision.

Recent research highlights the vulnerabilities of Vision-Language-Action (VLA) models to multimodal adversarial attacks, as presented in the study 'When Alignment Fails: Multimodal Adversarial Attacks on Vision-Language-Action Models.' The study introduces VLA-Fool, a framework that examines the adversarial robustness of VLAs under both white-box and black-box conditions, focusing on cross-modal misalignment that impacts decision-making.

When Alignment Fails: Multimodal Adversarial Attacks on Vision-Language-Action Models

One More Thing in AI – Your Shortcut to AI Mastery

When Alignment Fails: Multimodal Adversarial Attacks on Vision-Language-Action Models

Was this article worth reading? Share it

One More Thing in AI

LucidQuery AI

The Visualizer

Https

LangWatch

Synthesia

Ready to build your own newsroom?