HKMA unveils ART framework and a shared AI training data platform to harden Hong Kong banking

The Hong Kong Monetary Authority announced a coordinated package of measures this week to guide banks on safer, more accountable use of artificial intelligence. The initiative, branded ART for Accountability, Resilience, and Trust, combines updated supervisory guidance, extended sandbox testing of agentic AI, and a new AI training data sharing platform to be developed with Cyberport and the University of Hong Kong.

What the ART package does

At its core, ART is a supervisory and practical response to the rapid rise of generative and agentic AI systems inside financial services. The HKMA has issued refreshed supervisory guidance that incorporates lessons from its Generative A.I. Sandbox program. The guidance emphasises clear assignment of responsibility for AI actions, stronger validation and testing regimes, and operational safeguards that let banks intervene or suspend automated agents when they behave outside acceptable boundaries.

Complementing the guidance, the HKMA plans to expand sandbox work into a GenA.I. Sandbox++ aimed at agentic use cases where systems can plan and undertake multi step tasks. The authority says these trials will help shape how intervention controls and emergency suspension mechanisms should operate when AI agents are deployed in real banking workflows.

Training data platform aims to tackle a core weakness

Perhaps the most concrete operational measure in ART is the announcement of an A.I. training data sharing platform being built in collaboration with Hong Kong Cyberport Management Company Limited and the University of Hong Kong. The HKMA framed data quality assurance as a core design principle for the platform. Regulators and practitioners have increasingly noted that model validation and safety depend on consistent, well documented datasets for training and testing. The platform is intended to provide banks and their vetted technology partners with access to curated datasets and shared validation tools to improve reproducibility of tests and comparability of model performance.

The authority has not published a launch date, access rules, or fine technical or governance details for the platform. Announcements so far indicate the initiative will prioritise data quality checks, provenance records and controlled access that protects customer privacy and preserves proprietary interests. The HKMA said the platform will support effective training and testing of AI systems used in banking, where both data sensitivity and regulatory consequences are high.

Why the move matters for Hong Kong banks and tech partners

Hong Kong has been positioning itself as a regional fintech and AI hub, with government programmes that expand computing infrastructure and fund AI research. Banks are running more AI trials and integrating more sophisticated models into operations from onboarding and fraud detection to customer service. The ART package signals a shift away from experimentation towards structured, supervised adoption with clearer expectations from the regulator.

For banks, the practical measures reduce two common sources of regulatory friction. First, the supervisory guidance clarifies what documentation, monitoring and intervention capabilities supervisors will expect when agentic systems make autonomous decisions. Second, a shared data platform could reduce duplication of costly data preparation and provide standardised benchmarks that regulators and banks can use to judge model safety and bias.

For technology vendors and research partners, the platform offers a controlled environment to test models on representative, quality assured datasets without exposing raw customer data. For academic teams, access may provide empirical workloads that better reflect banking realities and accelerate responsible model research targeting finance.

Outstanding questions and constraints

Important details remain unresolved. The HKMA has not yet released the platform governance, participant eligibility criteria, privacy protection architecture, or a timetable. Those elements will determine whether the platform meaningfully reduces friction and risk or merely centralises datasets without sufficient transparency on curation and access. Stakeholders will be watching whether the project uses synthetic datasets, deidentified real data, strict contractual protections, or a combination of methods to reconcile utility with privacy obligations.

Another open issue is cross border data flows. Hong Kong banks operate in multiple jurisdictions and increasingly rely on international model components and cloud services. The new platform will need clear rules on what testing and training may include when datasets or models could cross legal boundaries, and how results will be shared with overseas supervisors and partners.

How this fits with Hong Kong policy direction

The ART package arrives against a broader Hong Kong policy push to scale up AI skills, infrastructure and governance. The government has allocated funds and established research and computing resources to make the city attractive for AI development. Financial regulators have been piloting sandbox approaches and releasing model governance guidance in recent years. ART links these strands by pairing supervisory expectations with a practical infrastructure intervention, the training data platform.

That pairing reflects a wider regulatory approach seen in other financial centres. Supervisors there are moving from high level principles to concrete tools and testbeds, recognising that safe deployment of powerful AI models depends on both governance and accessible, high quality test data.

What to watch next

Watch for the HKMA to publish the supervisory guidance document and for Cyberport and the University of Hong Kong to publish technical and governance specifications for the training data platform. Stakeholder consultations, participant onboarding rules, and a pilot timeline will be key signals of how quickly the platform will become operational and whether it can meaningfully raise model validation standards across Hong Kong banking.

In the months ahead, banks taking part in GenA.I. Sandbox++ trials will produce practical examples of agentic AI in regulated settings. How those pilots demonstrate intervention controls, explainability and incident response will inform final supervisory expectations and the platform's design.