What does Taiwan FSC AI Attack Recommendations say about AI in banking?

What are the key points of Taiwan FSC AI Attack Recommendations?

• The risk: frontier models raise the speed, scale and reliability of vulnerability discovery and weaponisation; the paper says time-to-exploit may shrink from days or weeks to hours, while many patches released at once (a 'patch flood') congest testing, validation and deployment at institutions and their suppliers.
• Zero trust: phishing-resistant multi-factor authentication, device health, network segmentation, connection control and continuous verification, plus stronger authentication, authorisation checks, anomaly detection and traffic control for back-end services and APIs.
• Monitoring and development: move security monitoring from alert visibility to early detection and rapid triage, and shift security left into development, testing and deployment, with CI/CD security controls, automated testing, signature verification, regression testing and rollback arrangements.
• Suppliers: make SBOM (software bill of materials) and dependency management routine, and progressively define key suppliers' patch SLAs, reporting deadlines, mitigations and alternatives.
• Joint defence: build on the F-ISAC information-sharing centre with clearer tiered reporting thresholds and deadlines for high-risk vulnerabilities, major supply-chain events and dense patch waves.
• Short-term readiness: identify priority services, critical systems and external exposure; take stock of patching capacity; triage vulnerabilities on risk; apply compensating controls such as isolation, privilege reduction, rate limits, temporary blocking rules, endpoint detection and response or segmentation where patching must wait; prepare for shutdown, degraded operation and external communication; and run resilience drills, with backups that include offline or immutable copies.
• Defensive AI: plan it into medium and long-term capability for vulnerability detection and patching, monitoring and incident response, while keeping human review, audit trails, deactivation and rollback mechanisms.
• International context: the paper reviews responses in the United States, the United Kingdom and Japan, including the short-term readiness measures issued jointly by Japan's FSA and the Bank of Japan, notes supervisory action in Singapore, Hong Kong, Australia and Korea, and cites a Five Eyes cyber agencies' statement that boards and senior management should treat cyber security as a core business risk, an approach the paper says matches what financial institutions need.

What did Taiwan FSC AI Attack Recommendations change for banks?

The paper applies the FSC's cyber-resilience policy to frontier AI models and moves the subject from technical defence to governance, operations and resilience led by senior management. It adds no new rulebook: its first five measures are mapped to the FSC's existing resilience blueprint and zero-trust reference guidance. The FSC says it will keep the direction under rolling review as model capabilities, international supervisory practice and industry needs develop.