Adversarial Robustness in LLM-Based Multi-Agent Trading Systems: A Systematic Vulnerability Analysis

Large Language Model (LLM)-based multi-agent trading systems have emerged as promising tools for autonomous market analysis, where specialized agents collaborate through structured communication to make trading decisions. However, the adversarial robustness of these systems remains largely underexplored, particularly at the inter-agent communication layer. In this work, we systematically evaluate multi-agent trading pipelines under realistic role-specific attack scenarios. For role-specific vulnerability analysis, we decompose the trading system into four functional roles—Analyst, Researcher, Trader, and Risk Manager—and design role-specific adversarial scenarios for each agent: Data Poisoning and Indirect Prompt Injection for Analysts, Persuasive Adversary attacks for Researchers, Objective Hijacking for Traders, and Jailbreaking for Risk Managers. Through extensive experiments on six major technology stocks, we evaluate both single-agent and combined attack scenarios. Our results reveal critical vulnerabilities across all agent roles, demonstrating that prompt-based attacks can effectively propagate through the collaborative decision pipeline. These findings underscore the urgent need for role-specific defense strategies and robust mechanisms in multi-agent trading architectures.