Financial Security Institute Develops Financial AI Safety Evaluation System... Full Operation to Begin Next Year

Korea’s First Evaluation Framework Specialized for Finance
10 Criteria Covering Hallucinations, Bias, and Security Threats
Briefing Session on the 14th and Demand Survey in September
Pilot Verification Scheduled for the Second Half of the Year

The Financial Security Institute has established an integrated evaluation system to comprehensively assess risks that may arise during financial companies’ use of artificial intelligence (AI), such as hallucinations, malfunctions, and security incidents. After conducting pilot verification with financial institutions this year, the organization plans to formally launch its evaluation work starting in 2027.

Sangwon Park, President of the Financial Security Institute, is being interviewed by The Asia Business Daily at the Financial Security Institute office in Yeouido, Seoul. 2026.6.18 Photo by Hyunmin Kim

On the 12th, the Financial Security Institute announced that it has developed Korea's first “Financial AI Safety and Reliability Evaluation Framework” specialized for the financial industry.

This framework was devised to systematically manage new risks stemming from increased AI adoption in the financial sector, especially as regulatory requirements like network separation have been relaxed. As the scope of AI use expands across core operations in financial companies, tailored evaluation standards for dealing with hallucinations, malfunctions, and security incidents have become necessary.

The Financial Security Institute established detailed criteria by analyzing domestic systems—including the Financial Services Commission’s “Financial Industry AI Guidelines,” the Financial Supervisory Service’s “Financial AI Risk Management Framework (RMF),” and the AI Basic Act—as well as international evaluation frameworks, such as the International Organization for Standardization’s “ISO/IEC 42001” (AI Management System Standard) and the UK AI Safety Institute’s evaluation tool “Inspect.”

The evaluation system consists of 10 items across two categories: AI reliability and AI safety. The reliability category includes model performance management, data quality management, fairness and bias, and explainability. The evaluation will check whether the model’s performance indicators and thresholds are appropriately set, if ongoing monitoring of hallucinations and performance degradation is in place, and more. It will also scrutinize the accuracy, completeness, and consistency of data, how bias is managed during model operation, whether AI use and decision-making processes are explained to end users, and if procedures for raising objections and seeking redress are established.

The safety category comprises six items: AI-specific security threats, detection and response to AI-specific attacks, protection and management of AI assets, validation of external models and data, scaling the security management system, and security verification and operational management.

This includes evaluating whether adversarial attacks are detected and blocked through input review and filtering, and whether security risks related to AI models, associated assets, and open-source components are managed. The evaluation also covers the safety of externally sourced models and data, supply chain security risks, systems preventing internal data leakage, and compliance with restrictions on cross-border data transfers.

The Financial Security Institute will hold an online briefing session for financial companies on August 14 and conduct a demand survey in September. In the second half of the year, the institute will examine the framework’s effectiveness through pilot verifications and refine the evaluation criteria. Starting next year, the evaluation work will first be carried out for member companies of the institute and, once stable, may be expanded to all financial institutions. The institute is also considering ways for this framework to be used as an AI safety and reliability verification and certification system in alignment with the AI Basic Act.

This content was produced with the assistance of AI translation services.