When Hallucination Costs Millions: Benchmarking AI Agents in High-Stakes Adversarial Financial Markets

AI models fail at detecting financial misinformation: Even state-of-the-art models achieved only 28% accuracy on financial decision-making tasks without tools, and plateaued at 67.4% accuracy even with unlimited access to professional resources—significantly below the 80% human baseline. This gap is particularly concerning given that $30 billion was lost to crypto exploits in 2024.

AI agents actively choose unreliable information sources: Models systematically preferred web search over authoritative specialized tools, even when correct answers were directly accessible. This suggests the problem isn't a knowledge gap but a fundamental limitation in how AI agents evaluate source reliability and resist SEO-optimized misinformation and social media manipulation.

Current AI evaluation methods don't capture real-world risks: Existing benchmarks measure task completion in controlled settings, but the paper demonstrates that AI agents face a "critical blind spot"—they cannot operate reliably in adversarial, high-stakes environments where misinformation is weaponized and errors are irreversible, like financial markets.

Trial-and-error metrics hide dangerous autonomous behavior: Pass@k metrics (which measure success across multiple attempts) mask problematic trial-and-error behavior that would be unacceptable in autonomous financial decision-making, where errors cannot be undone and have real financial consequences.

High-stakes deployment requires new safety standards: The findings suggest that before deploying AI agents in domains like finance, healthcare, or critical infrastructure, we need evaluation frameworks specifically designed to test resilience against active deception and irreversible consequences, not just task accuracy.