arXiv:2509.23694v5 Announce Type: replace-cross 
Abstract: Search agents connect LLMs to the Internet, enabling them to access broader and more up-to-date information. However, this also introduces a new threat surface: unreliable search results can mislead agents into producing unsafe outputs. Real-world incidents and our two in-the-wild observations show that such failures can occur in practice. To study this threat systematically, we propose SafeSearch, an automated red-teaming framework that is scalable, cost-efficient, and lightweight, enabling sandboxed safety evaluation of search agents. Using this, we generate 300 test cases spanning five risk categories (e.g., misinformation and prompt injection) and evaluate three search agent scaffolds across 17 representative LLMs. Our results reveal substantial vulnerabilities in LLM-based search agents, with the highest ASR reaching 90.5% for GPT-4.1-mini in a search-workflow setting. Moreover, we find that common defenses, such as reminder prompting, offer limited protection. Overall, SafeSearch provides a practical way to measure and improve the safety of LLM-based search agents.

تقديم SafeSearch، وهو إطار عمل آلي لاختبار أمان وكلاء البحث المعتمدين على نماذج اللغة الكبيرة (LLM)، يسلط الضوء على الثغرات المرتبطة بربط LLM بالإنترنت. يتيح هذا الإطار إجراء اختبارات منهجية لوكلاء البحث عبر فئات مخاطر متعددة، مما يكشف عن نقاط ضعف كبيرة في مخرجاتهم.

La introducción de SafeSearch, un marco de red-teaming automatizado para evaluar agentes de búsqueda basados en modelos de lenguaje de gran tamaño (LLM), destaca las vulnerabilidades asociadas con la conexión de los LLM a Internet. Este marco permite pruebas sistemáticas de los agentes de búsqueda contra varias categorías de riesgo, revelando debilidades significativas en sus resultados.

L'introduction de SafeSearch, un cadre d'évaluation automatisé pour les agents de recherche basés sur des modèles de langage de grande taille (LLM), met en lumière les vulnérabilités associées à la connexion des LLM à Internet. Ce cadre permet des tests systématiques des agents de recherche contre diverses catégories de risques, révélant des faiblesses significatives dans leurs résultats.

The introduction of SafeSearch, an automated red-teaming framework for evaluating LLM-based search agents, highlights the vulnerabilities associated with connecting large language models (LLMs) to the Internet. This framework enables systematic testing of search agents against various risk categories, revealing significant weaknesses in their outputs.