When AI Meets Wall Street: A Survey on Trustworthy AI in Fintech

Financial AI systems face unique vulnerabilities across their entire lifecycle: The paper identifies that attacks on AI in finance are particularly dangerous because they operate within "continuously operated financial AI pipelines" where small algorithmic changes can "amplify into persistent, system-level financial harm"—meaning a minor manipulation could cascade into widespread financial damage affecting many people.

Seventeen distinct types of attacks threaten financial AI systems: The research catalogs specific attack methods including data poisoning (corrupting training information), model poisoning (manipulating the AI itself), adversarial attacks on decision boundaries, prompt injection in AI chatbots, and deepfake-based fraud in identity verification—each requiring different defensive strategies tailored to finance rather than generic AI security approaches.

Finance-specific constraints make standard AI security inadequate: Unlike other industries, financial AI must contend with "accounting plausibility" (attacks must hide within normal financial records), federated data across institutions, and continuous retraining—meaning existing adversarial machine learning research doesn't fully address how attacks persist and spread in financial systems.

Automation amplifies the consequences of AI failures: The paper emphasizes that the "automation-amplified downstream effects" of compromised financial AI mean that attacks aren't limited to individual transactions but can trigger system-wide consequences, making trustworthiness in financial AI a systemic risk issue rather than just a technical one.

Current financial institutions lack adequate stress-testing for AI attacks: The research identifies a critical gap: there is no established framework for "lifecycle-aware stress testing" of financial AI systems, leaving institutions unprepared to identify and defend against sophisticated attacks before they cause real financial harm.