What does MAS Circular MAS/TCRS/2024/05 say about AI in banking?

What are the key points of MAS Circular MAS/TCRS/2024/05?

• Section 2.1 covers deepfakes and GenAI-enabled phishing, citing incidents in 2023-2024 including a deepfake video-conference fraud in Hong Kong and deepfakes used to defeat facial biometric authentication; mitigation includes liveness detection, awareness campaigns, additional verification for high-risk transactions and deepfake scenarios in incident response.
• Section 2.2 covers malware generation and enhancement, including polymorphic malware; mitigation includes a multi-layered cyber defence strategy and consideration of AI tools to detect polymorphic malware.
• Section 3.1 covers data leakage from GenAI deployments: user policies and awareness, security best practice for in-house models, due diligence on third-party or open-source GenAI solutions, and data loss prevention and firewalls.
• Section 3.2 covers GenAI model and output manipulation, including data poisoning; mitigation includes model and data governance, access controls to training data and foundation models, maker-checker and human-in-the-loop controls, continuous monitoring and validation, contingency in business continuity plans and information sharing.
• Section 4 concludes and the Appendix summarises each threat's impact and countermeasures by people, process and technology.
• The circular's description says MAS planned a follow-up information paper on AI model risk management for 4Q 2024 or 1Q 2025 (issued 5 December 2024) and points FIs to Project MindForge's industry-led whitepaper on emerging GenAI risks and opportunities.

What did MAS Circular MAS/TCRS/2024/05 change for banks?

It was the first MAS paper to catalogue generative-AI-specific cyber threats for financial institutions, separating the attacker's use of GenAI from risks in the FI's own GenAI deployments. It has been followed by MAS's September 2025 paper on deepfakes and by MAS's 2026 work on frontier-AI cyber threats, and it is the cyber companion to MAS's model risk and governance papers.

What does MAS/TCRS/2024/05 say banks should do about generative AI cyber risks?

MAS circular MAS/TCRS/2024/05, 'Cyber Risks Associated with Generative Artificial Intelligence', of 30 July 2024 is an awareness paper rather than a rulebook, but it lists the mitigation measures MAS expects financial institutions to consider. For GenAI-enabled attacks they include liveness detection against deepfakes, user awareness campaigns, extra verification for high-risk transactions and privileged roles, deepfake scenarios in incident response and a multi-layered cyber defence. For GenAI deployments they include user policies, data classification, due diligence on third-party and open-source models, data loss prevention, access controls on training data and foundation models, human-in-the-loop and maker-checker controls, continuous monitoring and validation, and contingency plans in business continuity planning. Banks should map these against their existing technology risk controls.

The paper separates threats by who is using the GenAI. Where attackers use it (deepfakes, phishing, malware), the mitigations are largely strengthened conventional controls. Where the FI itself deploys GenAI, the risks are data leakage and manipulation of the model or its outputs, and the controls move toward data governance, access control and model oversight, which overlap with the model risk expectations in MAS's December 2024 information paper.

It is the earliest in a series. MAS's September 2025 deepfakes paper (MAS/TCRS/2025/06) deepens section 2.1, and the November 2025 consultation paper says risks from external actors' use of AI, such as AI-powered cyber-attacks, are outside the scope of the proposed Guidelines and covered by these cyber papers instead.

WHAT THIS MEANS IN PRACTICE

• Test facial-recognition onboarding and authentication against deepfake injection, including liveness detection.
• Add deepfake and GenAI-phishing scenarios to staff awareness programmes and incident response playbooks.
• Write a staff policy on what data may be entered into public GenAI tools and enforce it with data loss prevention.
• Run third-party and open-source GenAI models through due diligence before use.
• Restrict access to training data and foundation models and require maker-checker approval for changes.
• Include GenAI services in business continuity plans.