Hong Kong regulators push "AI vs. AI" to police AI risks with GenA.I. Sandbox++
Hong Kong's financial regulators have launched the Generative Artificial Intelligence (GenA.I.) Sandbox++, a citywide push to scale safe, compliant AI across finance. The Hong Kong Monetary Authority, Securities and Futures Commission, Insurance Authority, and Mandatory Provident Fund Schemes Authority announced the initiative on 5 March, in collaboration with Hong Kong Cyberport.
It builds on the 2024 GenA.I. Sandbox and now spans banking, securities and capital markets, asset and wealth management, insurance, MPF, and stored value facilities. The focus stays tight: risk management, anti-fraud, and customer experience - with an explicit "AI vs. AI" approach to manage AI-related risks.
What's new - and why it matters
- Regulatory guidance and technical support while you test AI in a controlled environment.
- Complimentary GPU access at Cyberport's AI Supercomputing Centre to accelerate prototyping and validation.
- Cross-sector collaboration so banks, insurers, asset managers, and fintechs can co-develop practical solutions.
Regulators signaled this is part of a longer-term push under "Fintech 2030" to support responsible innovation. Translation: build, but build with guardrails and measurable controls.
Priority use cases regulators want to see
- Insurance: AI-driven underwriting, triage, and claims processing with auditable decision trails.
- Markets and distribution: compliance assessments embedded in product recommendation and suitability workflows.
- Pensions: tooling to improve MPF servicing, inquiry handling, and operational efficiency.
- Customer experience: intelligent chatbots with clear fallback paths and disclosure controls.
- Fraud and financial crime: pattern detection, anomaly alerts, and "AI vs. AI" defenses against synthetic fraud.
How financial institutions can get value fast
- Start with one high-friction process where risk is controllable and data is available (e.g., KYC remediation, claims intake, call-center triage).
- Define model risk controls upfront: data lineage, prompt logging, testing thresholds, human-in-the-loop, and kill-switches.
- Run red-team exercises against generative models to probe prompt injection, data leakage, and hallucination failure modes.
- Quantify ROI with hard metrics: handling time, loss avoidance, false positive/negative rates, and compliance exceptions.
- Align with second-line teams early: legal, compliance, model risk, cyber, and data privacy.
What participants get inside the sandbox
- Supervisory feedback while iterating - reducing uncertainty on what will pass production scrutiny.
- Technical support to evaluate architectures, safeguards, and monitoring design.
- A safer place to test sector-specific and cross-sector AI applications before full rollout.
Eddie Yue, chief executive of the Hong Kong Monetary Authority, said the launch of GenA.I. Sandbox++ sits within the authority's Fintech 2030 strategy to support responsible innovation in financial services. The signal is clear: bring your AI projects into a framework that proves control, not just capability.
Next steps
- Map two to three candidate use cases to the sandbox's focus areas (risk, anti-fraud, CX) and shortlist one to pilot.
- Prepare a minimal but complete control stack: data governance, access controls, content filters, monitoring, and incident playbooks.
- Line up a cross-functional squad (business owner, data science, engineering, compliance, model risk) and define week-by-week deliverables.
For background on the regulators and infrastructure involved, see the Hong Kong Monetary Authority site HKMA and Cyberport's AI Supercomputing Centre.
If you're exploring sector-specific playbooks and training, start here: AI for Finance and AI for Insurance.
Your membership also unlocks:
