Singapore Regulatory Crosswalk
Verification required. Claims on this page about Singapore law were compiled on 7 October 2026 from secondary web sources and from memory of the law. On 10 October 2026, rows SG5 and SG6 were checked against the MAS Guidelines on Artificial Intelligence Risk Management of 7 October 2026 and are tagged [Verified]. No other primary text was read, and every other claim is tagged [Reported], [Recalled], [Conflict] or [Unverified] as explained on the section home page. Check the primary text before relying on any of them. This is not legal advice.
One row per instrument in the Singapore AI Regulatory Hub, with the wiki controls that relate to it and its verification status. The last column holds the primary URL and the date it was retrieved, for the rows that have been checked. A note in brackets says when only part of the row was checked. The MAS Guidelines are also mapped paragraph by paragraph further down this page.
The same data is available as a CSV file for GRC tooling: crosswalk.csv.
Legal force sorts each row under one of seven labels, so that law is not mistaken for guidance or for a voluntary framework. The label is this wiki's reading of the sources and carries the same verification caveat as the rest of the row.
| Label | Meaning |
|---|---|
| Binding law | Enacted law or regulation that applies generally in the jurisdiction, including provisions that start on a later date. |
| Binding in scope | Binding only on a sector, a free zone, a state or province, or public bodies. |
| Guidance | Published by a government or regulator, and not binding in itself. |
| Voluntary | A standard, code or framework that an organization chooses to adopt. |
| Proposed | A bill, draft or consultation, or a proposal that lapsed. It is not law. |
| Context | A policy, authority, programme or summary row that sets no rule. |
| Unclear | The sources used do not settle whether it binds. |
| ID | Instrument | Legal force | Type / status | Wiki controls | Verification tag | To verify | Primary URL and retrieval date |
|---|---|---|---|---|---|---|---|
| SG1 | Model AI Governance Framework (2019; later editions) | Voluntary | Voluntary | AI-CTRL-012; AI-CTRL-037; AI-CTRL-039 | Recalled | Current edition | |
| SG2 | Model AI Governance Framework for Agentic AI (IMDA, 22 Jan 2026) | Voluntary | Voluntary | AI-CTRL-023; AI-CTRL-025; AI-CTRL-037; AI-CTRL-005 | Reported | Full text | IMDA: Model AI Governance Framework for Agentic AI, retrieved 10 Oct 2026 (address as cited in the MAS Guidelines, footnote 11; the document was not read) |
| SG3 | AI Verify testing framework and toolkit; Global AI Assurance Pilot (Feb 2025) | Voluntary | Voluntary testing and assurance | AI-CTRL-038; AI-CTRL-039 | Reported | Principle list (11, from memory) | |
| SG4 | PDPA; PDPC Advisory Guidelines on use of Personal Data in AI Recommendation and Decision Systems, and on use of Personal Data in Generative AI | Binding law | Binding law; guidance | AI-CTRL-013; AI-CTRL-007; AI-CTRL-017 | Reported | Guideline content; exceptions for research or business improvement | PDPC: AI Recommendation and Decision Systems; PDPC: Generative AI, retrieved 10 Oct 2026 (titles and addresses as cited in the MAS Guidelines, footnote 30; the PDPC pages were not read) |
| SG5 | MAS Guidelines on Artificial Intelligence Risk Management (7 Oct 2026) | Guidance | Issued 7 Oct 2026; sets out supervisory expectations. Takes effect 7 Oct 2027 for Sections 3 and 4, and 7 Oct 2028 for Sections 5 and 6 | AI-CTRL-001; AI-CTRL-005; AI-CTRL-006; AI-CTRL-008; AI-CTRL-009; AI-CTRL-010; AI-CTRL-011; AI-CTRL-012; AI-CTRL-013; AI-CTRL-014; AI-CTRL-016; AI-CTRL-017; AI-CTRL-023; AI-CTRL-024; AI-CTRL-025; AI-CTRL-028; AI-CTRL-029; AI-CTRL-030; AI-CTRL-031; AI-CTRL-032; AI-CTRL-033; AI-CTRL-034; AI-CTRL-035; AI-CTRL-037; AI-CTRL-038; AI-CTRL-039; AI-CTRL-040; AI-CTRL-041; AI-CTRL-044; AI-CTRL-045 | Verified | Sector notices and other MAS guidelines that the Guidelines refer to | MAS: Guidelines on Artificial Intelligence Risk Management for Financial Institutions, retrieved 10 Oct 2026 |
| SG6 | MAS FEAT principles and Veritas | Voluntary | Voluntary financial-sector principles; the MAS Guidelines of 7 Oct 2026 say they continue to apply (para 1.2) | AI-CTRL-041 | Verified | Veritas status | MAS: FEAT principles; MAS Guidelines, para 1.2, retrieved 10 Oct 2026 (status checked in the MAS Guidelines; the FEAT document itself was not read) |
| SG7 | CSA Guidelines and Companion Guide on Securing AI Systems (2024) and agentic addendum (17 Jun 2026) | Voluntary | Voluntary | AI-CTRL-005; AI-CTRL-008; AI-CTRL-025 | Reported | Text | |
| SG8 | Cybersecurity Act (critical information infrastructure) | Binding in scope | Binding | AI-CTRL-005; AI-CTRL-035 | Recalled | Scope |
MAS Guidelines on AI Risk Management: paragraph crosswalk
MAS issued the Guidelines on Artificial Intelligence Risk Management on 7 October 2026. They set out MAS's supervisory expectations for financial institutions and replace the proposal that MAS consulted on in 2025. The table maps each group of paragraphs to wiki controls, test cases and templates. It was written from the published Guidelines, read on 10 October 2026, and the mapping is this wiki's own judgment.
| Point | What the Guidelines say | Paragraph |
|---|---|---|
| Who it applies to | All financial institutions, in proportion to their size and risk profile. For locally incorporated institutions under consolidated supervision, or that own critical information infrastructure, it applies on a group basis | 1.2, 2.1 |
| What counts as AI | Machine-based systems or models that derive outputs through learned premises. Tools whose outputs come only from predefined logic or rules are not AI for this purpose. Generative AI and AI agents are in scope | 1.3, 1.5 |
| When it takes effect | 7 October 2027 for Sections 3 and 4 (oversight; identification, inventory and risk materiality). 7 October 2028 for Sections 5 and 6 (life cycle controls; capability and capacity) | 1.8 |
| Lighter route | Basic governance policies and procedures are enough where poor performance or unavailability of the AI is unlikely to have a material adverse impact | 2.3 to 2.5 |
| Related instruments | The FEAT principles continue to apply. Institutions may refer to the CSA guidelines on securing AI systems, the IMDA Model AI Governance Framework and its agentic AI version, and AI Verify | 1.2, footnotes 3 and 11 |
Coverage shows how far the wiki covers the expectation: Direct (a wiki control addresses it) or Partial (the last column says what is missing).
| Paragraphs | Expectation | Wiki controls | Test cases and templates | Coverage | What the wiki does not cover |
|---|---|---|---|---|---|
| 1.6, 2.1 to 2.2 | Apply the Guidelines in proportion to the institution's size and activities and to the risk materiality of each AI use case | AI-CTRL-011, AI-CTRL-037 | TC-L01-005, TC-L01-009, SG-T2 | Direct | |
| 2.3 to 2.5 | Basic governance where an AI failure is unlikely to have a material impact: a named senior manager, permitted and prohibited uses, human review of outputs, an approved list of AI tools, staff education, compliance checks and periodic review | AI-CTRL-012, AI-CTRL-010, AI-CTRL-001, AI-CTRL-044 | TC-L02-001, TC-L02-023, TC-L04-001, TC-L03-038 | Direct | |
| 3.1 to 3.4 | The board approves and reviews the governance approach for AI risk, addresses AI in the risk appetite framework and understands AI well enough to challenge | AI-CTRL-012, AI-CTRL-037 | TC-L02-005, TC-L02-009, TC-L02-024 | Partial | No control or test case covers board understanding of AI or quantitative risk appetite measures |
| 3.5 | Senior management implements the framework, assigns roles across business lines and control functions, escalates AI incidents and threshold breaches, and resources AI risk management | AI-CTRL-012, AI-CTRL-035, AI-CTRL-044 | TC-L02-004, TC-L01-003, TC-L17-018 | Direct | |
| 1.5, 3.6 | Local senior management stays accountable where a group framework is used | AI-CTRL-012 | TC-L02-004 | Partial | The wiki does not address group and local accountability |
| 4.2 to 4.4 | Identify AI use consistently across the institution, including AI inside third-party services and shadow AI, under a designated control function | AI-CTRL-001, AI-CTRL-011 | TC-L01-008, TC-L04-001, TC-L05-002, TC-L06-003 | Direct | |
| 4.5 to 4.9 | Keep an accurate AI inventory with key attributes, linked to other inventories, with agent-specific attributes where relevant | AI-CTRL-011, AI-CTRL-006 | TC-L01-001, TC-L01-004, TC-L06-002, SG-T2 | Direct | |
| 4.10 to 4.13 | Assess the risk materiality of each use case on impact, complexity and reliance, before and after controls, with a control function approving the rating | AI-CTRL-011, AI-CTRL-037 | TC-L01-005, TC-L01-009, TC-L02-007, TC-L02-009, SG-T2 | Direct | |
| 5.3 | Contingency plans and fallback for high-risk use cases, and tested activation of kill switches | AI-CTRL-040, AI-CTRL-025 | TC-L01-018, TC-L06-022, TC-L06-023 | Direct | |
| 5.4 | Data management: fitness for purpose, representativeness, quality, classification, security, privacy, and auditability and lineage | AI-CTRL-017, AI-CTRL-029, AI-CTRL-013 | TC-L13-001, TC-L13-002, TC-L13-006, TC-L15-012, SG-T3 | Direct | |
| 5.5 to 5.6 | Transparency and explainability matched to the use case and its audience, with customer notice and redress for high-impact decisions | AI-CTRL-041, AI-CTRL-039 | TC-L03-003, TC-L03-029, TC-L03-008 | Direct | |
| 5.7 to 5.8 | Define fair outcomes, run fairness assessments that consider protected attributes and their proxies, and document mitigation | AI-CTRL-041 | TC-L03-031, TC-L03-032 | Direct | |
| 5.9 | Human oversight: assigned roles, competent overseers with authority to intervene, design that supports oversight, and review of interventions and near misses | AI-CTRL-045, AI-CTRL-023 | TC-L01-012, TC-L03-040, TC-L03-041 | Direct | |
| 5.10 to 5.11 | Third-party AI: test it in the institution's own context, secure contractual visibility of changes, assess the supply chain and concentration risk, plan for failure and keep a right to audit | AI-CTRL-014, AI-CTRL-031, AI-CTRL-040 | TC-L16-008, TC-L16-010, TC-L16-011, TC-L16-012, TC-L16-023, TC-L16-025, TC-L03-042 | Direct | |
| 5.12 to 5.13 | Document why an AI system, model or feature was selected over simpler alternatives | AI-CTRL-033 | TC-L01-002 | Partial | No control covers the rationale for selecting a model or feature |
| 5.14 to 5.15 | Evaluation and testing against measurable thresholds, including stress, adversarial and sub-population tests; guardrails for known limitations; failure modes of generative AI and AI agents | AI-CTRL-034, AI-CTRL-039, AI-CTRL-041 | TC-L12-016, TC-L12-017, TC-L13-010, TC-D08-006, TC-D08-019 | Partial | Overfitting mitigation and statistical validation methods are outside the wiki |
| 5.16 | Technology and cybersecurity controls: secured environments, access control and separation of duties, and governance of third-party components | AI-CTRL-032, AI-CTRL-016, AI-CTRL-005, AI-CTRL-024 | TC-L15-004, TC-L15-006, TC-L12-013, TC-L13-024, TC-L16-015 | Direct | |
| 5.17 | Document development so that an independent party can understand and reproduce it | AI-CTRL-030, AI-CTRL-038 | TC-L13-013, TC-L14-019 | Direct | |
| 5.18 to 5.21 | Review before deployment by people not involved in development, formal independent validation for high-risk use cases, and findings reported to the approval body | AI-CTRL-033, AI-CTRL-038 | TC-L01-002, TC-L02-013, TC-L14-002 | Partial | The wiki gates release on a security review. It has no control on independent model validation |
| 5.22 | Technology and cybersecurity review before go-live: secure design, vulnerability assessment, penetration testing, red teaming and tests against adversarial threats | AI-CTRL-034, AI-CTRL-033, AI-CTRL-032 | TC-D08-002, TC-D08-020, TC-L14-002 | Direct | |
| 5.23 | Ongoing monitoring with metrics and thresholds, drift checks, AI incident and issue management, user feedback, and logging of prompts and responses for generative AI and agents | AI-CTRL-009, AI-CTRL-035, AI-CTRL-008 | TC-L14-016, TC-L17-018, TC-L06-028, TC-D12-001 | Direct | |
| 5.24 | Re-validation and review, regular for high-risk use cases and when triggered by changes, alerts or new risks | AI-CTRL-034, AI-CTRL-038 | TC-L12-018, TC-L01-015, TC-D08-021 | Direct | |
| 5.25 | Change management: a way to judge how significant a change is, re-approval, version control with rollback, and tighter controls for automatic updates | AI-CTRL-033, AI-CTRL-030 | TC-L06-034, TC-L14-012, TC-L14-013, TC-L01-015 | Direct | |
| 5.26 | Retirement and decommissioning of AI | AI-CTRL-011, AI-CTRL-028 | TC-L01-019, TC-L12-024, TC-L06-035 | Direct | |
| 6.1 to 6.2 | Competence and conduct of the people who develop, deploy and use AI, with regular review of capability and capacity | AI-CTRL-044 | TC-L03-038, TC-L02-022 | Direct | |
| 6.3 | Technology infrastructure adequate for performance, scalability and resilience | AI-CTRL-032, AI-CTRL-040 | TC-L15-002, TC-L15-019 | Direct |
Of the 26 groups of paragraphs, 21 are a direct match and 5 are partial. The partial ones are where the Guidelines go into model risk management and board governance, which sit outside a security and governance wiki: board understanding of AI and risk appetite measures, accountability between a group and its Singapore entity, the rationale for selecting a model, statistical validation methods and independent model validation.
The same mapping is available as a CSV file: mas_ai_risk_guidelines.csv.