Shilling attackers often operate within or across social platforms, leveraging their influence to manipulate collective opinions, distort reputation metrics, and skew product visibility. Shilling attack detection in recommender systems remains a persistent challenge due to the increasing sophistication of adversarial profiles. Traditional detection pipelines, including statistical and supervised models, often fail to generalize across multimodal attacks that simultaneously target ratings, reviews, and user behaviour (i.e., both ratings and reviews). This paper presents LLM-MAD, a novel multi-agent detection framework that leverages prompt-driven Large Language Models (LLMs) for robust and generalizable detection of adversarial shilling attacks. Our framework, by dissecting behavioral signals across various modalities, models the implicit and explicit trust dynamics and the adversarial influence pathways that are central to social networks. Our proposed model deploys three specialized GPT-based agents, each focusing individually on reviews, ratings, and profile behaviour. The outputs from these agents are then fused using a meta-agent, which a rule-based orchestrator governs to issue final judgments and justifications. To ensure robustness across different adversarial strategies, we simulate three distinct attack scenarios (ratings-only, reviews-only, and profile-level) on constructed infected datasets, reflecting realistic obfuscation tactics. Evaluated on both the infected Amazon and Yelp datasets, LLM-MAD achieves high accuracy in tackling multi-modal shilling attacks. Experimental results demonstrate that our model outperforms both classical baselines and advanced hybrid models, achieving classification accuracies of 98% and 84.2% on the Amazon and Yelp datasets, respectively, while remaining cost-effective in deployment. Our work highlights the strength of LLM-driven agent collaboration in building resilient and transparent recommender systems under multi-modal adversarial conditions.