arXiv:2606.16821v2 Announce Type: replace 
Abstract: Large language model (LLM)-based search agents synthesize open-web content into actionable recommendations on behalf of users, creating a risk that attacker-published pages are transformed into endorsed claims. We introduce SearchGEO, a controlled evaluation framework for measuring endorsement corruption in LLM-based web-search agents, combining a web-evidence manipulation pipeline, a five-mode attack taxonomy, and multiple output-level metrics. We evaluate 13 LLM backends on 308 cases each. Results show that vulnerability patterns vary across backends: overall attack success rate (ASR) ranges from 0.0% on Claude-Sonnet-4.6 to 31.4% on Gemini-3-Flash, the strongest attack mode differs by model family, and the same deployment scaffold could amplify or decrease ASR on different backends. An auxiliary agent-skill probe, where endorsement becomes an install command, exposes a sharp split among otherwise robust backends: Claude over-rejects while GPT over-trusts. These findings argue for treating recommendation reliability under adversarial search content as a first-class dimension of backend safety evaluation.

قدمت دراسة حديثة إطار عمل يسمى SearchGEO، مصممًا لتقييم ضعف وكلاء البحث المعتمدين على نماذج اللغة الكبيرة (LLM) تجاه فساد التأييد الناتج عن التلاعب بمحتوى الويب. قامت الدراسة بتقييم 13 خلفية LLM، كاشفة عن معدلات نجاح هجوم متباينة، حيث سجل Gemini-3-Flash أعلى معدل بنسبة 31.4% وClaude-Sonnet-4.6 أدنى معدل بنسبة 0.0%.

Un estudio reciente presentó SearchGEO, un marco diseñado para evaluar la vulnerabilidad de los agentes de búsqueda basados en modelos de lenguaje (LLM) a la corrupción de respaldo causada por la manipulación de contenido web. La investigación evaluó 13 backends de LLM, revelando tasas de éxito de ataque variables, siendo Gemini-3-Flash el más alto con un 31.4% y Claude-Sonnet-4.6 el más bajo con un 0.0%.

Une étude récente a introduit SearchGEO, un cadre conçu pour évaluer la vulnérabilité des agents de recherche basés sur des modèles de langage (LLM) à la corruption d'approbation causée par la manipulation de contenu web. La recherche a évalué 13 backends LLM, révélant des taux de succès d'attaque variables, avec Gemini-3-Flash affichant le plus élevé à 31,4 % et Claude-Sonnet-4.6 le plus bas à 0,0 %.

A recent study introduced SearchGEO, a framework designed to evaluate the vulnerability of large language model (LLM)-based search agents to endorsement corruption caused by web content manipulation. The research assessed 13 LLM backends, revealing varying attack success rates, with Gemini-3-Flash showing the highest at 31.4% and Claude-Sonnet-4.6 the lowest at 0.0%.