arXiv:2510.24034v1 Announce Type: new 
Abstract: Despite rapid advancements in text-to-image (T2I) models, their safety mechanisms are vulnerable to adversarial prompts, which maliciously generate unsafe images. Current red-teaming methods for proactively assessing such vulnerabilities usually require white-box access to T2I models, and rely on inefficient per-prompt optimization, as well as inevitably generate semantically meaningless prompts easily blocked by filters. In this paper, we propose APT (AutoPrompT), a black-box framework that leverages large language models (LLMs) to automatically generate human-readable adversarial suffixes for benign prompts. We first introduce an alternating optimization-finetuning pipeline between adversarial suffix optimization and fine-tuning the LLM utilizing the optimized suffix. Furthermore, we integrates a dual-evasion strategy in optimization phase, enabling the bypass of both perplexity-based filter and blacklist word filter: (1) we constrain the LLM generating human-readable prompts through an auxiliary LLM perplexity scoring, which starkly contrasts with prior token-level gibberish, and (2) we also introduce banned-token penalties to suppress the explicit generation of banned-tokens in blacklist. Extensive experiments demonstrate the excellent red-teaming performance of our human-readable, filter-resistant adversarial prompts, as well as superior zero-shot transferability which enables instant adaptation to unseen prompts and exposes critical vulnerabilities even in commercial APIs (e.g., Leonardo.Ai.).

تقدم ورقة جديدة AutoPrompt، وهي طريقة لأتمتة اختبار أمان نماذج النص إلى صورة لتعزيز سلامتها ضد المطالبات العدائية. هذا مهم لأنه يعالج نقاط الضعف في هذه النماذج، التي يمكن استغلالها لتوليد صور غير آمنة. من خلال تحسين كفاءة اختبار هذه النماذج دون الحاجة إلى الوصول المباشر، يمكن أن يؤدي AutoPrompt إلى تطبيقات ذكاء اصطناعي أكثر أمانًا في المجالات الإبداعية.

Un nuevo artículo presenta AutoPrompt, un método para automatizar el red-teaming de modelos de texto a imagen para mejorar su seguridad contra los prompts adversariales. Esto es significativo porque aborda las vulnerabilidades de estos modelos, que pueden ser explotados para generar imágenes inseguras. Al mejorar la eficiencia de las pruebas de estos modelos sin necesidad de acceso directo, AutoPrompt podría llevar a aplicaciones de IA más seguras en campos creativos.

Un nouvel article présente AutoPrompt, une méthode d'automatisation du red-teaming des modèles de texte à image pour améliorer leur sécurité contre les invites adversariales. Cela est important car cela aborde les vulnérabilités de ces modèles, qui peuvent être exploités pour générer des images dangereuses. En améliorant l'efficacité des tests de ces modèles sans nécessiter d'accès direct, AutoPrompt pourrait conduire à des applications d'IA plus sûres dans les domaines créatifs.

A new paper introduces AutoPrompt, a method for automating the red-teaming of text-to-image models to enhance their safety against adversarial prompts. This is significant because it addresses the vulnerabilities of these models, which can be exploited to generate unsafe images. By improving the efficiency of testing these models without needing direct access, AutoPrompt could lead to safer AI applications in creative fields.

AutoPrompt: Automated Red-Teaming of Text-to-Image Models via LLM-Driven Adversarial Prompts

arXiv:2511.18684v1 Announce Type: new 
Abstract: Robust concept removal for text-to-image (T2I) and text-to-video (T2V) models is essential for their safe deployment. Existing methods, however, suffer from costly retraining, inference overhead, or vulnerability to adversarial attacks. Crucially, they rarely model the latent semantic overlap between the target erase concept and surrounding content -- causing collateral damage post-erasure -- and even fewer methods work reliably across both T2I and T2V domains. We introduce Instant Concept Erasure (ICE), a training-free, modality-agnostic, one-shot weight modification approach that achieves precise, persistent unlearning with zero overhead. ICE defines erase and preserve subspaces using anisotropic energy-weighted scaling, then explicitly regularises against their intersection using a unique, closed-form overlap projector. We pose a convex and Lipschitz-bounded Spectral Unlearning Objective, balancing erasure fidelity and intersection preservation, that admits a stable and unique analytical solution. This solution defines a dissociation operator that is translated to the model's text-conditioning layers, making the edit permanent and runtime-free. Across targeted removals of artistic styles, objects, identities, and explicit content, ICE efficiently achieves strong erasure with improved robustness to red-teaming, all while causing only minimal degradation of original generative abilities in both T2I and T2V models.

قدم الباحثون تقنية Instant Concept Erasure (ICE)، وهي نهج جديد لإزالة المفاهيم بشكل موثوق في نماذج النص إلى صورة (T2I) والنص إلى فيديو (T2V). تلغي هذه الطريقة الحاجة إلى إعادة التدريب المكلفة وتقلل من عبء الاستدلال، بينما تعالج الثغرات الأمنية ضد الهجمات العدائية. تستخدم ICE تقنية تعديل الوزن في خطوة واحدة دون تدريب، مما يضمن إلغاء تعلم دقيق ودائم دون أضرار جانبية للمحتوى المحيط.

Investigadores han presentado el Instant Concept Erasure (ICE), un enfoque novedoso para la eliminación robusta de conceptos en modelos de texto-a-imagen (T2I) y texto-a-video (T2V). Este método elimina la necesidad de reentrenamiento costoso y minimiza la sobrecarga de inferencia, al tiempo que aborda las vulnerabilidades a ataques adversariales. ICE emplea una técnica de modificación de pesos en una sola vez, sin necesidad de entrenamiento, que asegura un desaprendizaje preciso y persistente sin daños colaterales al contenido circundante.

Des chercheurs ont introduit l'Instant Concept Erasure (ICE), une nouvelle approche pour la suppression robuste de concepts dans les modèles de texte-à-image (T2I) et de texte-à-vidéo (T2V). Cette méthode élimine le besoin de réentraînement coûteux et minimise la surcharge d'inférence tout en abordant les vulnérabilités aux attaques adversariales. ICE utilise une technique de modification de poids en une seule fois, sans formation, qui garantit un désapprentissage précis et persistant sans dommages collatéraux au contenu environnant.

Researchers have introduced Instant Concept Erasure (ICE), a novel approach for robust concept removal in text-to-image (T2I) and text-to-video (T2V) models. This method eliminates the need for costly retraining and minimizes inference overhead while addressing vulnerabilities to adversarial attacks. ICE employs a training-free, one-shot weight modification technique that ensures precise and persistent unlearning without collateral damage to surrounding content.

AutoPrompt: Automated Red-Teaming of Text-to-Image Models via LLM-Driven Adversarial Prompts

Was this article worth reading? Share it

Blunge

Grubby.AI

Prompt Builder