Commentary · Technology disputes · Cyber incident response
MAS AI risk management guidelines: when “model glitches” become regulatory failures
- MAS issued its final Guidelines on Artificial Intelligence Risk Management on 7 October 2026. They take effect on 7 October 2027.
- They apply to all financial institutions and all forms of AI, including generative AI and AI agents, with controls scaled to how material each use is.
- For digital asset businesses, trading bots, AI risk engines, on-chain surveillance and automated KYC now need the same governance as payments, AML/CFT and cyber.
- Firms stay accountable for third-party AI they can't fully inspect.
Singapore has moved from "responsible AI" as a slogan to responsible AI as a supervisory expectation. That shift will bite hardest where AI, trading and digital assets meet.
What MAS has done
MAS consulted on proposed guidelines in November 2025 and issued the final Guidelines on Artificial Intelligence Risk Management on 7 October 2026. They build on the earlier FEAT principles (Fairness, Ethics, Accountability and Transparency), but go much further, into concrete frameworks, controls and documentation. In outline, they expect:
- Board and senior management oversight of AI risk, integrated into existing enterprise risk frameworks.
- Identification and inventories of AI use across the firm, with each use assessed for risk materiality.
- Life cycle controls covering data, testing and evaluation, monitoring, change management and human oversight.
- Capabilities and capacity: the people, skills and infrastructure to manage AI safely.
The approach is proportionate. Where poor performance or unavailability of an AI tool is unlikely to have a material impact, basic governance may be enough. Material uses need the full suite. Firms also remain accountable for AI supplied by vendors, and can't rely on a supplier's self-attestation alone.
Why this matters for digital asset businesses
The interesting intersection is between three layers of regulation:
- the AI Guidelines;
- the existing MAS technology risk and cyber requirements; and
- the regimes for digital payment token services, digital token service providers and stablecoins under the Payment Services Act and the Financial Services and Markets Act.
Together they add a new dimension to compliance. Firms will have to show not just that their infrastructure is secure and compliant, but that the AI models running execution, risk, surveillance and customer touchpoints are governed and monitored to the same standard.
The use cases in scope
- Agentic trading agents and algorithmic market-making
- On-chain analytics for market surveillance
- Automated KYC and AML screening
- AI-driven customer support and credit or collateral assessment
Each needs an entry in the AI inventory, a risk materiality assessment, guardrails, and clear lines of responsibility tied into the firm's existing technology risk and conduct regimes.
When "model glitches" become regulatory failures
A trading bot that misfires, a risk engine that misprices collateral, or an agentic system that interacts with smart contracts in unexpected ways will no longer be treated as just a model error. Regulators will ask whether the firm treated it as a governed AI use case, with appropriate testing, oversight and accountability.
The same facts matter in disputes. When an AI-driven failure causes customer losses, the firm's AI inventory, risk assessments, testing records and human-oversight logs become key evidence on whether it acted with reasonable care. Firms that can't produce them will find both regulatory and civil exposure harder to defend. The October 2026 Binance incident, where an internal AI tool was used to move between engineers' computers, shows why. See lessons from the Binance incident.
What boards should do now
- Take AI out of the "innovation" silo. Bring it into the same control stack as payments, AML/CFT, cyber and operational resilience.
- Build the inventory and rate each use for materiality before the October 2027 effective date.
- Review vendor contracts for AI tools: audit rights, testing, certifications and incident notification.
- Keep the records (testing, monitoring and human oversight) that will be needed if something goes wrong.
Frequently asked questions
When do the guidelines take effect?
Do they cover AI bought from vendors?
Related: the Bitget exploit and third-party vendor liability, and whether exchanges can contract out of duties to users.
Developed from a LinkedIn post written during the MAS consultation, and updated for the final guidelines. General information only, not legal advice.
