Adversarial Attacks Threaten Multi-Agent LLM Trading Systems.

Multi-agent trading systems powered by Large Language Models (LLMs) are becoming increasingly sophisticated, with specialized agents collaborating through structured communication to make trading decisions. While this collaboration enhances effectiveness, it also introduces significant vulnerabilities. This research demonstrates that corrupted signals, introduced through accessible means like source data or agent prompts, can propagate through the system and result in real financial losses. The study is the first to empirically analyze how adversarial signals infiltrate these systems and their impact on final decisions within the financial domain. It decomposes a typical trading pipeline into roles like Analyst, Researcher, Trader, and Risk Manager, applying role-specific attacks. Furthermore, it evaluates four communication topologies under data and agent-level attacks, using a novel "Adversarial Signal Preservation Score" to understand design robustness. A key finding is that no single architecture is inherently immune to these threats, providing crucial insights for developing more secure agentic trading systems.