The Financial Security Institute has built a new evaluation system aimed at catching the hazards that come with financial firms deploying artificial intelligence, including hallucinations, system malfunctions and security breaches.
According to Asia Business Daily, the institute confirmed on the 12th that it had completed Korea’s first dedicated evaluation framework for AI reliability and safety within financial services, describing it as a response to the growing use of AI in core banking functions following the loosening of network separation rules.
The push comes as financial institutions expand AI use across critical operations, prompting the institute to argue that bespoke evaluation standards are now needed to manage the fallout from model errors, unreliable outputs and cyber incidents.
To build the framework, the institute drew on a mix of domestic and international standards, including the Financial Services Commission’s AI guidelines for the financial industry, the Financial Supervisory Service’s AI risk management framework, and South Korea’s AI Basic Act, alongside overseas benchmarks such as ISO/IEC 42001, the international standard for AI management systems, and Inspect, the evaluation tool developed by the UK’s AI Safety Institute.
The system is structured around ten criteria split across two pillars: reliability and safety. On the reliability side, assessors will examine model performance management, data quality, fairness and bias, and explainability, checking whether performance thresholds are properly calibrated, whether hallucinations and performance decay are continuously monitored, and whether the data feeding these models is accurate, complete and consistent.
This pillar also covers how bias is controlled during live operation, whether customers are given clear explanations of AI-driven decisions, and whether channels exist for people to challenge or seek redress for those decisions.
The safety pillar spans six areas: threats specific to AI systems, detection and response to AI-targeted attacks, protection of AI assets, vetting of external models and data, scalability of security governance, and ongoing security verification. In practice, this means testing whether adversarial attacks are caught and blocked through input screening, whether risks tied to AI models, related assets and open-source components are properly managed, and whether externally sourced models and data are vetted for safety.
It also takes in supply chain security, safeguards against internal data leakage, and adherence to rules governing cross-border data transfers.
On the rollout timeline, the Financial Security Institute plans to run an online briefing for financial firms on 14 August, followed by a demand survey in September. Pilot testing will follow in the second half of the year to refine the criteria before full evaluations begin in 2027, starting with the institute’s own member companies before a possible extension to the wider financial sector.
The institute is also weighing whether the framework could eventually double as a formal certification system for AI safety and reliability under the AI Basic Act.
Copyright © 2026 FinTech Global
