AI Model Risk Management for Banks in 2026: What APRA, MAS, BNM, HKMA and the CBUAE Now Expect

APRA's April 2026 letter: a step-change is required

On 30 April 2026, APRA wrote to all regulated entities after a targeted review of large banks, insurers and superannuation trustees. Its letter on artificial intelligence calls for a step-change, because governance, risk management, assurance and resilience are not keeping pace with adoption. Its main findings:

• Boards are still building the technical literacy to challenge AI risk, and rely too much on vendor presentations.
• Lifecycle controls are weak after deployment. Few entities continuously validate models for drift, bias or failure, and APRA says point-in-time, sample-based assurance is ill suited to probabilistic models.
• Cyber and identity controls have not adjusted to non-human actors such as AI agents, and staff use of AI tools outside approved controls relies on policy rather than technical prevention.
• Third parties are concentrated. Exit plans are rarely tested, and contracts often lack audit rights and model-update or incident notification clauses. APRA expects entities to map the full AI supply chain, including foundation model providers as fourth parties.

APRA is not proposing new requirements at this stage. It expects existing prudential standards to be met and has said it will take stronger supervisory action where AI risk is poorly managed. In practice that means AI now sits inside CPS 230 operational risk and CPS 234 information security reviews.

What each regulator has said

Two points stand out. First, Singapore's guidelines are still a consultation, so no effective date should be assumed. Second, the new US guidance deliberately leaves generative and agentic AI out of scope while the agencies consult further, which means the Asian and Gulf regulators are, for once, ahead of Washington on the hardest part of the problem.

Where regulators agree

Read side by side, the documents share five expectations. They are the backbone of any AI model risk programme in the region.

1. A complete AI inventory. APRA, MAS, the CBUAE and SR 26-2 all require one, covering in-house models, vendor models and AI embedded in purchased software.
2. Risk-tiering. Controls proportionate to materiality. MAS proposes impact, complexity and reliance as the dimensions; the CBUAE allows fully automated decisions only for low-risk, non-material processes.
3. Validation that continues after launch. Pre-deployment testing is no longer enough. Drift, bias and failure monitoring must run while the model is in use.
4. Human oversight matched to risk. Human involvement for high-risk decisions, and a route for customers to seek human review.
5. Third-party accountability. Outsourcing a model does not outsource the risk. Contracts, testing rights and exit plans must cover AI providers, including the foundation model behind a vendor product.

Why classic model validation struggles with generative AI

Traditional validation tests conceptual soundness, benchmarks outcomes and monitors performance against a stable model. Generative AI challenges each step. Outputs vary for the same input. The model may be a vendor service that changes without notice. There is often no clean ground truth to measure against. In a September 2026 speech, the chair of the BIS Financial Stability Institute said model risk expectations built for transparent statistical models must be revisited for large language models.

Until that happens, banks are adapting with practical measures:

• Scenario and red-team testing of prompts, including prompt injection and data leakage attempts.
• Output sampling and scoring in production, with thresholds that trigger review.
• Version pinning and change notice clauses, so a vendor model update is treated as a model change.
• Guardrails and fallbacks for any AI that supports a critical operation, so the service continues if the model is withdrawn.

Fraud is where the risk is most visible. In early 2024, an employee in engineering firm Arup's Hong Kong office joined a video call with deepfaked senior colleagues and made transfers totalling about HK$200 million, as CNN reported. Payment approval controls now need to assume that voices and faces can be faked.

A practical AI model risk framework

This lifecycle maps onto ISO/IEC 42001, which gives the management system a certifiable structure. For the governance picture beyond banking, see our guides to AI governance in Australia and New Zealand, AI governance in Saudi Arabia, the UAE and Pakistan, and governing AI agents under ISO/IEC 42001.

Running AI model risk in GRCLens

GRCLens holds the AI inventory as a register linked to risks, controls and suppliers, so a vendor model appears both in the model register and in third-party risk. Materiality assessments, validation reports and monitoring evidence are attached to each model with their dates, and the same records map to APRA prudential standards, MAS TRM, BNM RMiT and ISO/IEC 42001. Because GRCLens can run inside your own infrastructure, with its AI evidence review hosted locally, model documentation never leaves the bank's environment.

Frequently asked questions

How Security Solution Consultants can help

Security Solution Consultants helps banks, insurers and financial institutions build AI inventories, risk-tiering methods and validation programmes that meet APRA, MAS, BNM and Gulf expectations. See our compliance and risk management advisory and our AI governance guide for businesses. For the security testing side, read securing AI solutions. GRCLens then keeps the inventory, assessments and monitoring evidence current. Request a demonstration.