TradeTrap is a unified evaluation framework designed to systematically stress-test LLM-based autonomous trading agents under adversarial conditions. The paper decomposes trading agents into four core components—market intelligence, strategy formulation, portfolio and ledger handling, and trade execution—and injects controlled perturbations (data fabrication, MCP tool hijacking, prompt injection, memory poisoning, state tampering) into each component. Experiments are conducted in a closed-loop historical backtesting environment using real U.S. equity market data (NASDAQ-100) with identical initial conditions, enabling fair and reproducible comparisons across two agent architectures: adaptive (tool-calling) and procedural (pipeline-driven).

The results reveal that small, localized perturbations can propagate through the decision pipeline, leading to extreme concentration, runaway exposure, and large portfolio drawdowns. Adaptive agents are highly opportunistic but vulnerable to information-channel attacks (e.g., fake news and MCP hijacking), which degrade risk-adjusted performance and induce strategic paralysis. Procedural agents exhibit greater resistance to noisy inputs but are catastrophically affected by state or memory corruption, such as memory poisoning causing persistent underperformance and state tampering converting them into single-asset strategies or accumulating massive short positions. The study highlights that current LLM-based trading agents can be systematically misled at the system level, and the choice between agent architectures involves a trade-off between upside capture and the attack surface one is willing to harden. TradeTrap provides a reproducible benchmark with nine quantitative metrics and open-source code to facilitate future robustness evaluations.