FinRED: An Expert-Guided Benchmark Generation and Evaluation Framework for Financial LLM Red-Teaming

As large language models (LLMs) are increasingly deployed in high-stakes financial environments, existing safety benchmarks often fail to capture domain-specific risks, focusing instead on generic adversarial scenarios or simple disclaimer-based checks. This paper asks: how can we systematically generate and evaluate LLM resilience against nuanced, expertise-driven financial threats such as regulatory evasion and complex fraud?

The authors introduce FinRED, an applied red-teaming framework developed in collaboration with financial security experts. The framework operates through three core stages:

1. Taxonomy and Schema Definition: A two-level risk taxonomy maps global standards (e.g., FATF, EU DORA) to specific financial threats. Expert-defined JSON schemas provide structured templates for generating realistic threat scenarios.
2. Contextual Retrieval: The system retrieves relevant information from a curated corpus of financial regulations, audit reports, and security guidelines to ground the generated threats in real-world operational environments.
3. Seed Generation: Using these schemas and retrieved contexts, the framework generates "Behavioral Prompts" (seeds) that simulate sophisticated adversarial attacks. These seeds are then validated by experts for plausibility and realism.

FinRED provides a scalable, expert-validated pipeline that significantly outperforms generic benchmarks in identifying domain-specific vulnerabilities. By utilizing an expert-aligned evaluation rubric, the framework reduces critical false negatives in threat detection compared to static, one-size-fits-all rubrics. The framework has been successfully deployed in the Financial Security Institute (FSI) regulatory sandbox in South Korea, proving its practical utility for verifying the security of generative AI in real-world financial services.