Bank of Russia Issues Methodological Recommendations for AI Security in Financial Sector: Overview and Practical Steps

The Bank of Russia issued Methodological Recommendations No. 3-MR on June 16, 2026, addressing information security in AI development and use in the financial market. The document targets banks, non-credit financial organizations, payment system participants, and other market players. It builds on the Code of Ethics for AI in finance and introduces AI-specific terminology such as hallucinations, data drift, prompt injection, and poisoned datasets, aligning with national standards (GOST R). Risks are categorized into six types: data management, confidentiality, model malfunction, explainability, vendor/open source risks, and operational reliability. Threat models must follow FSTEC methodology; measures are proportional to risk severity. The AI lifecycle is divided into four stages: data preparation, development, training/testing, and operation. Specific threats include evasion, poisoning, model theft, and adversarial attacks. The document emphasizes supply chain security, requiring organizations to assess trust in external data, models, and open source components, and to use sanitized or synthetic data when sharing with vendors. Human validation is recommended for critical automated processes such as payments. Organizations should inventory AI components, assess risks, build threat models, implement security measures across lifecycle stages, and formalize an AI security policy. Red team testing, data minimization, output labeling, and emergency shutdown plans are highlighted. The recommendations are advisory but signal the regulator's direction, urging proactive compliance.