Learn the stack in the order it makes sense.

IMDA: frameworks and testing ecosystem

IMDA writes the Model AI Governance Framework family, know it as the framework agency, not a lawmaker.

PDPC: binding data protection

PDPC administers the only binding economy-wide law in this stack, the PDPA.

MAS: financial-sector regulation

MAS issues binding notices and supervisory expectations for financial institutions.

CSA: AI security guidance

CSA owns the security layer, guidelines, companion guide, and the agentic addendum.

Anchor instrument: MGF 2.0

Nearly every later instrument builds on or references this 2020 framework.

AI Verify: the assurance layer

Government-built testing tooling connects the frameworks to evidence.

Read the framework overview

Start with what the framework is for: structured overview of agentic risks and best practices for deployers.

Agent anatomy: the model

The LLM/SLM/MLLM "brain", governance starts by knowing what you are governing.

Agent anatomy: memory

Memory makes agents stateful, and makes data governance materially harder.

Multi-agent patterns

Sequential, supervisor, swarm, the pattern determines where accountability can attach.

Five risk types

Erroneous, unauthorised, biased actions, data breaches, disruption, your risk-register skeleton.

Dimension 1: assess and bound risks upfront

Risk assessment plus design-time limits is the framework's first line of defence.

Dimension 2: human accountability

Approval checkpoints and audited oversight, adapted human-in-the-loop for the agentic era.

Track v1.5

The May 2026 update added case studies and multi-agent risk guidance, a living document to watch.

PDPA baseline

The binding Act, obligations apply to AI exactly as to any other processing.

Consent

The default gateway for collection, use and disclosure.

Exceptions without consent

First and Second Schedules, where training data can lawfully come from without consent.

2024 AI advisory

PDPC's interpretation for ML systems, business improvement and research exceptions operationalised.

Anonymisation guidance

Anonymised data is outside the PDPA but carries re-identification risk, the guidance sets the test.

2026 GenAI finals

The July 2026 guidelines cover development, deployment and post-deployment for GenAI.

General notices are insufficient

AI-specific notification is the new bar for training consent.

Access and correction for GenAI

Best practices for honouring individual requests against trained models, a hard problem with official expectations.

Binding: TRM Notices

Statutory technology risk requirements, the floor for any FI deploying AI.

RTO of 4 hours

Recovery objectives constrain how critical AI services must be architected.

FEAT Principles

Fourteen principles that define MAS's fairness/ethics/accountability/transparency expectations.

FEAT 1: no systematic disadvantage

The fairness principle your models will be tested against first.

Veritas

The methodology and open-source toolkit for assessing FEAT conformance.

Proposed AI Risk Management Guidelines

Consultation closed 31 Jan 2026, not yet final, plan against the proposal, label it as such.

MindForge toolkit

The industry handbook aligned to the proposal, your implementation bridge.

ABS GenAI guardrails

Banking-specific GenAI guardrails that fed the MAS toolkit.

AI Verify framework

Eleven internationally accepted principles, technical tests plus process checks.

AI Verify toolkit

Run the open-source toolkit; the report is your primary assurance artefact.

Moonshot

LLM benchmarking and red teaming to operationalise the GenAI testing dimensions.

Global AI Assurance Pilot

Where Singapore codified testing norms with international testers and deployers.

Crosswalk: NIST AI RMF

The official October 2023 mapping, GOVERN/MAP/MEASURE/MANAGE to AI Verify.

Crosswalk: ISO/IEC 42001

The June 2024 mapping of Annex A controls to AI Verify.

SS ISO/IEC 42001

The certifiable standard, Annex ZA expressly names AI Verify as a voluntary alignment tool.

MGF-GenAI: testing and assurance dimension

The framework-level case for why third-party testing and common standards matter.

Governance charter

Adapt internal governance structures and measures, the dossier's first chapter.

Self-assess with ISAGO

Run the official self-assessment to baseline your alignment with MGF 2.0.

Risk-rating method

Use the severity-probability matrix to justify human-involvement decisions per use case.

Responsibility allocation

Chains of accountability across the value chain, name owners for every system.

Documented assessments

DPIA-style written records for data-source decisions, the artefact regulators ask for.

Threat modelling record

Security threat modelling and taint tracing outputs belong in the dossier.

AI inventory

A maintained inventory of AI systems and capabilities, every regime expects one.

Evidence: process checks

Documentary evidence validating processes, the difference between claimed and demonstrated governance.

Start with what does not exist

No comprehensive federal AI law. Everything else follows from that.

The White House sets policy by executive order

Orders direct the executive branch. They are not statutes.

OMB turns orders into agency requirements

Numbered memoranda, binding on agencies.

NIST writes the voluntary framework

Influential, and not law.

The FTC is the closest thing to a general AI regulator

Section 5, no AI statute required.

And the states carry most of the binding law

Which is why there is a separate Atlas for them.

FTC Act Section 5

Unfair or deceptive practices, technology neutral for over a century.

Title VII

Discriminatory selection procedures, whoever or whatever produced the score.

The ADA

Screen out risk and reasonable accommodation.

ECOA and Regulation B

Specific and accurate principal reasons for adverse action.

The TAKE IT DOWN Act

The one federal statute written for AI generated content.

EO 14148 clears the board

Revokes EO 14110.

EO 14179 sets the replacement policy

And directs the action plan and the OMB rewrite.

The AI Action Plan lays out the programme

Three pillars, many recommendations, zero obligations.

EO 14319 reaches private vendors

Through procurement, which is the only route an order has.

EO 14365 turns on the states

Litigation task force, funding conditions, preemptive standards.

EO 14409 is the current frontier model order

Security first, voluntary, and expressly not a licensing regime.

M-25-22 governs the buy

Read it before you write the proposal.

M-25-21 governs the use

Your buyer must be able to answer for your system.

High impact AI carries minimum practices

And a directive to stop if they cannot be met.

LLMs carry two extra contract terms

Truth seeking and ideological neutrality.

Your deployment becomes public

Annual inventory with risk determinations.

Start where the binding law is

The states.

The litigation route

DOJ task force, announced January 2026.

The naming route

A Commerce list of onerous state laws, due March 2026, not published as of this dataset's verification date.

The money route

Broadband funding conditioned on not enforcing conflicting AI laws.

The deception route

A state law requiring altered outputs could itself breach the FTC Act.

And the legislative route

The March 2026 framework, which Congress has not enacted.

CAISI evaluates frontier models

Voluntary, agreement based.

The AI Agent Standards Initiative

Launched February 2026.

Covered frontier models under EO 14409

Classified benchmarking, voluntary access.

The explicit bar on licensing

Read it before assuming a regime exists.

CISA guidance on securing AI systems

The voluntary security baseline.