Two financial centres, one month, one idea: know your agent. How a small state that cannot win the race for superintelligence chose to sell the one thing the racers cannot make for themselves.
Between 2018 and November 2026, Singapore built one of the most complete AI governance stacks in the world without passing an AI law: first principles, then tests, then testers, then runtime controls for agents. Hong Kong took a different road towards the same destination. This is the story of how Singapore became a trust haven in the arms race for superintelligence, what the Hong Kong mirror reveals about it, and what the haven still lacks. Facts are stamped as of 20 November 2026.
The Two-Minute Version
- The race for superintelligence is being run by two superpowers that do not trust each other and, increasingly, do not want global rules. A small open economy cannot win that race. Singapore decided instead to sell what the racers cannot produce for themselves: trust that someone else can check.
- It did this without an AI statute. From 2018 to 2021 it wrote principles (FEAT, the Model AI Governance Framework, the first National AI Strategy). From 2022 to 2024 it built instruments (AI Verify, the AI Verify Foundation, Project Moonshot, the generative AI framework). 2025 was the testing year, with real deployers, independent testers and real failures. 2026 moved to agents and runtime: the agentic framework at Davos, SAFR, the MAS AI risk guidelines, accredited testers and, on 19 November, a know-your-agent protocol.
- Hong Kong is the mirror, not the rival lead. It reached the agent question by a sandbox-and-supervise route backed by free state compute, and on one point it is ahead: its 2 November discussion paper floated a central registry of material agents, a kill switch and a “Know Your Agent” approach, seventeen days before MAS spoke.
- Singapore’s strength is method: open tests, independent testers, published runtime designs. Its weakness is authority. Outside the regulated sectors, nobody can stop a deployment.
- The critics are right that voluntary assurance can drift into box-ticking or capture, and that testing finished applications misses risks that appear during frontier training. A trust haven has to answer them, not out-market them.
- What the haven needs next: an enforceability backstop, mutual recognition of test results, runtime evidence rather than paperwork, agent identity standards that interoperate with Hong Kong’s registry, and the political will to set the trajectory for the Fork.
Prologue: One Month, Two Cities
It was the best of times for anyone selling agents, and a rather more anxious time for anyone asked to supervise them. On the morning of 19 November 2026, in a hall at Singapore EXPO, the Managing Director of the Monetary Authority of Singapore, Chia Der Jiun, gave the second-day keynote of the Singapore FinTech Festival. Much of it was about tokenised settlement. The passage that will matter longest was about software that acts on our behalf.
Under what he called a new phase of SAFR, the Safeguards for Agentic Finance at Runtime published in July, MAS will facilitate the development of a know-your-agent protocol. The idea is simple to state and hard to build. When a customer’s agent knocks on a bank’s door, the protocol should tell the bank who the agent is and whom it represents, what authority it holds, what mandate it operates under, and what assurance exists about its provenance and reliability. Banks can then decide what each agent may do, through a dedicated secure channel, and verify identity and authority before a single dollar moves. Alongside it, an Agentic Money Sandbox with the new Future of Finance Institute will let agents manage corporate treasury and execute payments with tokenised deposits or stablecoins, inside mandates set by the customer and controls set by the bank.
He also did something regulators rarely do on a festival stage. He told consumers not to hand their personal agents the login details to their financial accounts, because the assurance is not yet there. And he left a question deliberately open: who bears responsibility when an agent acts wrongly, or beyond its instructions? His answer was a principle rather than a rule: “Autonomy has to be bounded. Actions and outcomes have to come with accountability.” The day before, Deputy Prime Minister Gan Kim Yong, who chairs MAS, had set out three National AI Missions for finance, the third of which is safe agentic transactions.
Seventeen days earlier, across the South China Sea, the Hong Kong Monetary Authority had opened its own FinTech Week with a discussion paper on agentic AI governance in banking. It proposed a risk-based “Know Your Agent” approach to establish an external agent’s identity, its principal and its delegated authority. It went further than Singapore on supervision, floating a central registry of material agents, which the HKMA itself might maintain, and an explicit kill switch to bring a misbehaving agent to a safe state. On the same day it launched Project SAGE, Securing Agentic Finance with Governance and Enablement, with a workstream devoted to agent identity and authentication.
Two cities, one month, the same idea, and two different routes to it. Hong Kong arrived through a regulator-run sandbox, free state compute and supervisory papers. Singapore arrived through eight years of principles, open testing tools, independent testers and a published runtime architecture. Neither invented the phrase; industry had been using it since at least 2025. What matters is not who was first but what each route tells us about the kind of trust each city is selling.
Those of us who spent 2025 in the testing rooms, watching agents and chatbots fail in ways no policy paper predicted, learned to distrust tidy stories. So this is not a puff piece. Singapore is the protagonist because its journey is the more instructive one. Hong Kong is the mirror, and mirrors are useful precisely because they show what you would rather not see.
The same three words, spoken seventeen days apart, mark two theories of how to make machines answerable.
I. The Backdrop: An Arms Race Nobody Can Referee
Start with the race, because the haven only makes sense against it. The United States and China are competing to build ever more capable AI, and the leaders of the frontier labs now speak openly of systems that could outperform people at most cognitive work within years. Demis Hassabis has said Chinese models may trail the American frontier by only months. Dario Amodei, in his January 2026 essay “The Adolescence of Technology”, described this as the most dangerous window, and backed chip export controls against a China he calls a high-tech surveillance state. Both superpowers frame AI as a matter of national survival. Neither wants to be bound by the other.
Helen Toner, the executive director of Georgetown’s Center for Security and Emerging Technology, put the danger in one line to the US Senate Judiciary Committee on 22 April 2026. After laying out strong evidence of Chinese “distillation attacks” on American labs, she still warned against a race mentality: “the winner of any AI race between the U.S. and China is the AI”. That is the sentence a trust haven is built on. If the racers cannot slow down, someone has to be able to check what they are fielding.
The multilateral system is not going to do the checking. At the India AI Impact Summit in New Delhi in February 2026, around ninety countries and organisations endorsed a declaration built on voluntary, non-binding frameworks. Michael Kratsios, director of the White House Office of Science and Technology Policy, said the United States “totally reject[s] global governance of AI”, while offering American technology as a route to sovereignty. China, for its part, proposed a World AI Cooperation Organization at Shanghai’s World AI Conference in 2025, with itself at the centre. A global referee with a whistle is not coming.
Into that gap step the middle powers, and Singapore is the smallest serious one. It cannot build frontier models at scale. It has no vast compute estate; it spent years under a moratorium on new data centres and remains land- and energy-constrained. On 18 November, Deputy Prime Minister Gan noted that new US Commerce rules now require licences to export or re-export advanced chips to five countries, Singapore among them. The city is a transhipment and data-centre hub sitting directly on the export-control fault line. If it tried to compete on capability, it would lose.
Its ministers know this and say so. At the Singapore Conference on AI in April 2025, Minister for Digital Development and Information Josephine Teo began from an admission: “We may not design or produce the most advanced AI chips.” What Singapore could do, she argued, was put governance on the same footing as adoption, and keep talking to everyone, including China through a bilateral digital policy dialogue.
Max Tegmark of MIT and the Future of Life Institute, who helped convene that conference, explained the logic to WIRED in May 2025. Singapore can bring the two AI superpowers into the same room because it “gets along well with both East and West”, and because it has no plans to build artificial general intelligence itself. It is not a contestant. It has a stake in the contestants talking. In a race where every builder distrusts every other, the scarcest commodity is a place both sides will accept as a fair test bench.
That is what I mean by a trust haven. Not a tax haven, where rules are lighter, but the reverse: a jurisdiction whose stamp is worth something because it is hard to get, open to inspection and accepted on both sides of a geopolitical divide. Switzerland sold neutrality and secrecy. Singapore is trying to sell neutrality and verifiability.
A small state cannot win the race for capability. It can try to become the place where capability is checked.
II. 2018 to 2021: The Principles Era
The story begins in finance, which is where Singapore’s regulatory muscle has always been. On 12 November 2018, MAS introduced the FEAT principles for Fairness, Ethics, Accountability and Transparency in the use of AI and data analytics, co-created with the industry. They were principles, not rules. But they were written by a regulator that licenses, inspects and can penalise the firms it addresses, and that difference would matter later.
Two months after FEAT, on 23 January 2019, the Personal Data Protection Commission and IMDA released the first Model AI Governance Framework, billed as Asia’s first and promising detailed, readily implementable guidance for the private sector. It rested on two ideas: AI-assisted decisions should be explainable, transparent and fair, and AI should be human-centric. It organised the work into four areas that any board could recognise: internal governance; the decision-making model and how much human involvement it needs; operations management; and customer relationships. Its most durable contribution was vocabulary. “Human-in-the-loop”, “human-over-the-loop” and “human-out-of-the-loop” became the grammar of a decade of governance debate.
A second edition followed at the World Economic Forum in Davos on 21 January 2020, with an Implementation and Self-Assessment Guide for Organisations developed with the WEF, and a compendium of real use cases. It added robustness, reproducibility and auditability, and a sharper account of when human involvement matters: the nature of the harm, whether it can be reversed, and whether oversight is operationally feasible. That last test, reversibility, would resurface six years later in Hong Kong’s agentic paper.
Meanwhile MAS turned principles into tools. Veritas, a consortium first announced in November 2019, worked through three phases to 2023, producing white papers on fairness metrics for credit scoring and marketing and two versions of an open-source toolkit, with membership growing from 25 to 32 institutions. Veritas mattered less for its outputs than for its habit. It taught Singapore’s regulators to govern by building something with industry and publishing it, rather than by decree.
The first National AI Strategy was unveiled by then Deputy Prime Minister Heng Swee Keat on 13 November 2019, with five national projects from freight planning to chronic disease prediction and border clearance, and a National AI Office set up the month before. Singapore later described itself as among the first countries to publish a national AI strategy. That is fair; Canada and China came earlier, in 2017.
What Singapore chose not to do in this period is as revealing as what it did. It did not legislate. As late as June 2023, IMDA’s Lee Wan Sie told CNBC: “We are currently not looking at regulating AI.” Critics would later call this light touch. From close quarters it looked more like a sequencing decision: learn what good practice looks like before freezing it in statute, and keep the existing sector regulators as the teeth. Whether that sequencing ever reaches its final step is the question this essay keeps returning to.
Singapore chose to test before it legislated, and has not yet legislated.
III. 2022 to 2024: Building the Instruments

The pivot came in May 2022, again at Davos. Josephine Teo launched AI Verify, which Singapore billed as the world’s first AI governance testing framework and toolkit, released as a minimum viable product developed by IMDA and the PDPC. Ten organisations piloted it, from global platforms to a national airline and two home-grown start-ups. The claim is Singapore’s own, though I know of no earlier government-built governance-testing toolkit.
The most important thing about AI Verify was its modesty. It did not set ethical standards. It did not certify that a system was safe. It let developers demonstrate, with technical tests and process checks, that their claims about a system were true, and later versions mapped those tests to eleven internationally recognised governance principles. A test you can run is worth more than a principle you can cite, and an open test can be inspected by the people it judges. That was a bet on verifiability over virtue, and everything since has built on it.
On 7 June 2023, Singapore open-sourced AI Verify and placed it under a new AI Verify Foundation, launched with seven premier members, IMDA alongside Aicadium, IBM, Microsoft, Google, Red Hat and Salesforce, and more than sixty general members. Handing the tool to a foundation was a signal aimed abroad. A government-owned test is a national product. An open, multi-member test is a candidate standard. In October the Foundation and IMDA announced a Generative AI Evaluation Sandbox they billed as first of its kind, and in November MAS published the first phase of Project MindForge on the risks and opportunities of generative AI for banks.
Then came the second national strategy. Launching NAIS 2.0 on 4 December 2023, Lawrence Wong, then Deputy Prime Minister, set a target of 15,000 AI practitioners and explicitly rejected both laissez-faire and strict upfront rules in favour of a “pragmatic balance”: sandboxes with guardrails, and rules updated over time. He flagged two external commitments, a forthcoming ASEAN guide and a partnership with the UK’s AI Safety Institute. The ASEAN Guide on AI Governance and Ethics was endorsed in Singapore on 2 February 2024, with Singapore driving it. In May 2024 the Digital Trust Centre at NTU was designated as Singapore’s AI Safety Institute and joined the international network, where Singapore co-leads the testing and evaluation track. In November 2024 Singapore and the UK signed a memorandum of cooperation on safety research and joint testing. Singapore was positioning itself as a peer in the Western assurance network, not a follower within it.
Generative AI forced the next instrument. On 30 May 2024 Singapore published the Model AI Governance Framework for Generative AI, with nine dimensions: accountability, data, trusted development and deployment, incident reporting, testing and assurance, security, content provenance, safety and alignment research, and AI for the public good. A day later Teo released Project Moonshot, an open-source toolkit combining benchmarking, red-teaming and testing baselines for both models and applications, which she described as one of the world’s first of its kind. In 2024 the Cyber Security Agency also published guidelines and a companion guide on securing AI systems, and the PDPC issued advisory guidelines on personal data in AI recommendation and decision systems.
There was one piece of hard law in this period, and it is telling. On 15 October 2024 Parliament banned election deepfakes that realistically depict a candidate saying or doing something they did not, whether favourable or not. Singapore legislates when the harm is specific, the target is clear and the democratic stakes are immediate. It does not legislate for “AI” in general.
In December 2024 MAS published an information paper on AI model risk management drawn from a thematic review of banks: inventories, materiality assessments, independent validation, monitoring. Read in hindsight, that paper is the skeleton of the guidelines MAS would issue two years later.
By the end of 2024 Singapore had tools, a foundation, a safety institute, a GenAI framework and a supervisor quietly collecting evidence. What it did not yet have was proof that any of this worked on real systems, in real organisations, under real pressure.
Instruments are promises. 2025 was the year Singapore had to show they could be kept.
IV. 2025: The Testing Year
On 10 February 2025, at the AI Action Summit in Paris, Teo announced the Global AI Assurance Pilot, run by IMDA and the AI Verify Foundation. The design was the important part. Rather than test foundation models in a lab, the pilot paired specialist assurance firms with organisations deploying real generative AI applications, and asked them to test those applications in context.
By the end of May the results were in. IMDA billed the exercise as the world’s first technical testing of real-world GenAI applications: 16 specialist testers paired with 17 deployers across ten industries, including finance, healthcare, human resources and the public sector. The main report reached a conclusion that sounds obvious and is routinely ignored: generative AI risk depends on context, on industry, use case, culture, language and organisation, so subject-matter experts must be involved throughout testing, not consulted at the end. A draft Testing Starter Kit followed, covering hallucination, undesirable content, data leakage and adversarial prompts.
Anyone who was in those rooms will remember how quickly the comfortable assumptions went. A chatbot that behaved impeccably in English would wander in another language. A guardrail that held on the first turn would bend on the fifth. The gap between what a vendor’s model card promised and what a deployed application actually did was rarely dramatic, and almost always present. Having watched these frameworks being stress-tested from the inside, I came away convinced that the unit of assurance is the deployment, not the model.
Between the launch of the pilot and its results, Singapore hosted a different kind of gathering. On 26 April 2025, alongside the ICLR conference, the Singapore Conference on AI brought together an expert committee that included Yoshua Bengio, Stuart Russell, Max Tegmark, Dawn Song and two senior Tsinghua figures, Xue Lan and Zhang Yaqin. Its output, the Singapore Consensus on Global AI Safety Research Priorities, published on 8 May, organised safety research into three areas: risk assessment, trustworthy development, and control after deployment. It argued that some safety work benefits everyone, as aviation safety data benefits rival aircraft makers. And, buried in its control agenda, it listed research into protocols for the “verification of AI agent identities while they are using web services”. Eighteen months before MAS announced a know-your-agent protocol, agent identity was already on the research agenda Singapore had convened.
The convening mattered as much as the content. American and Chinese scientists agreed a safety agenda on Singapore soil while their governments tightened export controls on each other. A trust haven sells exactly that: not the answer, but a room in which both sides will accept the question.
On 7 July 2025 the pilot became permanent as the Global AI Assurance Sandbox, and its scope widened to agentic AI and risks such as data leakage and prompt injection. Teo called the absence of AI testing a serious gap that had to be filled. The design choices are worth listing, because they define Singapore’s model. The sandbox tests the application layer, not foundation models. The Foundation matches deployers with independent testers, and the tester runs the tests, not the government. Testing happens in the participants’ own environments and results stay confidential to the two parties, but participants must publish a structured summary of architecture, risks, test design, thresholds and challenges. By late 2026 the Foundation’s page counted 30 applications across 12 geographies and 14 industries. As far as I know, it remains the only standing government programme matching independent testers with deployers across sectors and borders.
A case study in Thai
The single most instructive output of the sandbox, for my money, is a short case study from Thailand. KBTG, the technology arm of Kasikorn Bank, brought “Jingjai”, an internal assistant that answers employees’ questions about benefits and insurance in Thai, English, mixed Thai-English and Romanised Thai. The tester, Vulcan, ran 1,276 test cases spanning 11 threats, 29 attack techniques and four language formats.
The findings read like a syllabus for the next decade of AI assurance. Attacks in Thai succeeded slightly more often than attacks in English. Romanised Thai looked safer, but probably because the model was misunderstanding the input rather than resisting it. The assistant showed subtle bias against rural Thai profiles compared with Bangkok ones. In Thai it produced content that could be read as investment guidance naming securities. In Romanised Thai it began to reveal its retrieval parameters and tools. The testers also probed risks no Western benchmark would think to include, such as lèse-majesté and the Thai-Cambodian border dispute.
The conclusion is the important part: an assistant in a regulated setting needs confidentiality and content controls that hold in every language its users actually speak, and it needs independent multilingual testing to prove they do. Safety that holds only in English is not safety in Southeast Asia. A trust haven in a region of hundreds of languages is well placed to see this.
The year closed with the agent question moving to the front. In October the Cyber Security Agency released a draft addendum on securing agentic AI for consultation. In November Parliament passed the Online Safety (Relief and Accountability) Act, creating an Online Safety Commission. And on 13 November MAS consulted on Guidelines on AI Risk Management that would apply to every financial institution, proportionately, and explicitly cover generative AI and AI agents, while the second phase of MindForge published an AI risk management toolkit drawn up with banks, insurers, asset managers and technology firms.
Testing in 2025 taught a simple lesson: the risk lives in the deployment, in the language, in the context. So must the assurance.
V. 2026: From Frameworks to Runtime

If 2025 was the year of testing, 2026 was the year Singapore tried to govern what an agent does while it is doing it.
On 22 January, at Davos, IMDA launched the Model AI Governance Framework for Agentic AI, described officially as first of its kind. I wrote about it in February, and I still know of no earlier government framework dedicated to agentic AI. Its four dimensions are operational rather than aspirational: assess and bound the risks upfront by limiting an agent’s autonomy and its access to tools and data; make humans meaningfully accountable, with defined checkpoints for approval; implement technical controls across the lifecycle; and enable end users to carry their share of responsibility. It named the risks plainly: agents that change their environment, update a customer database or make a payment, and humans whose oversight decays into automation bias. A May update added case studies and guidance on multi-agent systems, third-party agents and automation bias. The same month the final Starter Kit for testing LLM-based applications was published.
February brought the money and the politics. Budget 2026 created a National AI Council chaired by Prime Minister Lawrence Wong and set National AI Missions in four sectors, finance among them. In April Singapore hosted the plenary of the international AI standards committee, ISO/IEC JTC 1/SC 42, the first in ASEAN, and proposed ISO/IEC 42119-8, a testing methodology for generative AI built on AI Verify, the Starter Kit and the sandbox. This is the trust-haven strategy at its clearest: turn a national method into an international standard, so that the haven’s stamp travels.
On 3 July MAS published SAFR, Safeguards for Agentic Finance at Runtime, an industry white paper developed with financial institutions and FinTechs. Its governance checkpoints “verify and record” an agent’s proposed actions before execution, with policy-bound execution, real-time validation, auditability and interoperability. As I described it in July, the design rests on four components: an agent identity, a controls repository, a disposition engine that routes each proposed action to execute, escalate, deny or observe, and a tamper-evident audit log. That is a gate before the action and a receipt after it, drawn by a central bank. It does not require anyone to build it. It shows them, in public, what one looks like.
Then the supervisor spoke with authority. On 7 October MAS issued its final Guidelines on AI Risk Management, applying to all financial institutions and all forms of AI. They require board and senior-management oversight, an AI inventory, a materiality assessment for each use, lifecycle controls from data governance to testing, human oversight, monitoring and change management, and they keep institutions accountable for AI bought from third parties, including the duty to limit, suspend or replace a service that exceeds risk appetite. They are proportionate: basic policies suffice where an AI failure would not materially affect the firm or its customers. Oversight, inventory and materiality expectations take effect on 7 October 2027; lifecycle controls and capabilities by 7 October 2028. On agents specifically, MAS asks firms to review their controls as use expands and will consult on further agentic guidance in 2027. I set out what changed from the consultation on the day.
Is this the most comprehensive AI-in-finance regime in the world? It is among the first dedicated AI risk-management guidelines from a financial supervisor to cover every institution and every form of AI, agents included, and on breadth I know of nothing more complete. But claims of primacy deserve care. Canada’s OSFI finalised a model-risk guideline covering AI in September 2025, and the HKMA issued updated supervisory guidance eight days after MAS. MAS itself claimed no first, which is to its credit.
A week after the guidelines, on 14 October, IMDA published an Agentic Testing Starter Kit, a four-step method (identify risks, design realistic tests, run tests and analyse logs, monitor and re-test) with metrics such as tool-use accuracy, handoff accuracy and the rate of human intervention. The same day the AI Tester Accreditation Programme, announced in May, went live with five firms accredited for twelve months by method and risk area. The Foundation is careful to say accreditation is not a guarantee that any AI system is safe, and that buyers keep responsibility for accepting risk. That candour matters. Accrediting testers is how a haven makes its stamp mean something; overselling the stamp is how it loses it.
The privacy and safety regulators moved in parallel. The Online Safety Commission began operating on 29 June. On 20 July the PDPC issued advisory guidelines on personal data in generative AI that place primary responsibility on the organisations deploying systems and flag agentic AI for particular attention. On 3 November the Future of Finance Institute began operations with three programmes: PathFin.ai for knowledge, BuildFin.ai for standards and common tools building on SAFR, and SafeFin.ai for applying the MAS guidelines, agents included. Guidance tells firms what to do. Shared infrastructure lowers the cost of doing it.
And so to the Singapore FinTech Festival. On 18 November DPM Gan set out three National AI Missions for financial services under the National AI Council. The first pools transaction data across five banks so that AI can spot scam patterns no single bank sees, with a design study for an industry-wide Scam Detection Utility from early 2027. The second uses AI and realistic synthetic data to widen access to finance for small firms and start-ups. The third is safe agentic transactions, with trials of purpose-bound money and common approaches to identity, authority, accountability and interoperability. His formulation was crisp: “Agents must operate within defined authority and limits.” Institutions, he added, remain accountable for actions taken on their behalf.
The next morning Chia announced the know-your-agent protocol and the Agentic Money Sandbox described in the prologue. The protocol is not a world first as a concept: industry frameworks predate it, Hong Kong floated the approach in a supervisory paper on 2 November, and MAS does not claim novelty. What may be new is a central bank committing to facilitate a common protocol for agent-to-institution information exchange, tied to a live sandbox and sitting inside a published runtime design. As far as I can find, no other central bank has done that, though I would not stake much on the word “first”.
Step back and the arc is plain: principles in 2018, tests in 2022, testers in 2025, runtime gates and agent identity in 2026. Each step was voluntary or supervisory, never statutory, and each was published for others to inspect, copy or criticise.
Singapore did not write a law for agents. It drew the gate, published the test and accredited the testers. Whether anyone must walk through the gate is another matter.
The sandbox state: governing by building together
Step back from the timeline and a method appears. Singapore does not govern AI mainly by writing rules and waiting for breaches. It governs by building shared places where the state, the companies that deploy AI and the firms that test it work on the same problem at the same table, and then it turns what they learn into guidance. Call it proactive governance: the regulator as convenor, toolmaker and co-designer, not only as referee.
AI Verify is the template. The state built a testing toolkit, then gave it away: open-sourced in June 2023 and placed in a foundation that companies could join and extend. The 2025 Global AI Assurance Pilot paired 17 deployers with 16 specialist testers on live applications, and the July 2025 sandbox made that pairing permanent. The output was not a certificate but a body of practice: what to test, how to test it, and what a test can and cannot tell you. Those of us who sat in those rooms learnt quickly that the value was less in any single score than in a deployer, a tester and a regulator arguing over the same failure until they agreed what “good” looked like.
Finance runs the same playbook one step further. MAS’s consortia, Veritas for fairness and MindForge for generative AI, wrote their toolkits with the banks rather than for them. In November 2026 that habit became an institution. The Future of Finance Institute, operational from 3 November and led on an interim basis by Alan Lim of MAS, starts with AI in finance through three programmes: PathFin.ai for knowledge and communities, BuildFin.ai for common standards and tools building on SAFR, and SafeFin.ai to help firms apply the new AI risk guidelines and extend the MindForge toolkit. Its first big project is the Agentic Money Sandbox, where banks set the controls and decide the terms on which they accept each agent’s instruction. The regulator supplies the protocol and the arena; the industry supplies the systems and the risk appetite; the rules come afterwards, written from evidence.
Malaysia shows the other serious route. Its AI Governance Bill, due to be tabled in early 2027 after a consultation that drew 317 submissions, is horizontal and risk-based. It imposes a statutory duty of care, creates a central AI authority with sector regulators in the lead, and requires risk assessment, incident reporting and corrective action from organisations deploying AI in Malaysia, wherever the system is hosted. That is something Singapore does not have: a power to compel. If an agent harms a customer outside the regulated sectors, Malaysia will be able to act, while Singapore will still be relying on goodwill, contracts and reputation.
The two approaches are not rivals so much as halves. Singapore has the method, the tested tools and the habit of building with industry; Malaysia will have the authority. A law without instruments risks becoming a compliance ritual, with duties nobody knows how to evidence. Instruments without a law risk becoming optional, used by the conscientious and ignored by the rest. The region needs both, and the sensible future is Malaysian duties discharged with Singapore-style evidence.
Then there is the third route, the most common of all: frameworks and recommendations that never leave the page. The ASEAN Guide on AI Governance and Ethics, expanded for generative AI in January 2025, is thoughtful and entirely voluntary. The Delhi Declaration from India’s AI Impact Summit in February 2026 was, by its own description, non-binding. Most national AI strategies in the world sit in this category: principles that everyone endorses and nobody operates. This is the paper-tiger problem I described in the companion piece on governance platforms. Singapore began in exactly this place in 2019. What separates it now is not that it legislated, because it has not, but that it kept building until its principles had test suites, accredited testers, a sandbox and, this week, a protocol.
| Route | Example | Strength | Weakness |
|---|---|---|---|
| Proactive, sandbox-led | Singapore (AI Verify, Global AI Assurance Sandbox, Future of Finance Institute) | Builds shared tools and evidence with industry; moves at the speed of the technology | No general power to stop a harmful deployment |
| Legislative | Malaysia (AI Governance Bill, 2027) | Duty of care, incident reporting, a central authority with teeth | Duties may outrun the means to evidence them |
| Supervisory sandbox | Hong Kong (HKMA GenAI Sandbox, joint regulator sandbox) | Regulator-run testing with free compute; fast agentic guidance in banking | Thinner legal backbone outside finance |
| Framework-only | ASEAN Guide; Delhi Declaration; most national strategies | Consensus and shared vocabulary | Nothing operational: the paper tiger |
A framework tells you what good looks like. A sandbox shows you whether you can build it. A law tells you what happens if you do not.
VI. The Other City: Hong Kong’s Route
Hong Kong deserves a fair hearing. It, too, has no horizontal AI statute and is not planning one. It, too, governs through sector regulators using existing powers. But where Singapore built testing infrastructure and handed it to a foundation, Hong Kong built a regulator-run sandbox on state compute and learned by supervising.
The foundations were laid in the same year as Singapore’s. In November 2019 the HKMA issued high-level principles on AI for banks and consumer-protection principles for big data and AI, first-generation texts that sit alongside MAS’s FEAT. In 2024 the pace picked up. The HKMA set out expectations for customer-facing generative AI in August, including the right to request human intervention, and in September told larger banks to consider AI for suspicious-activity monitoring, with board-signed plans due by March 2025, more directive about adoption than anything MAS has issued.
The other regulators followed. In November 2024 the SFC’s circular on generative AI told licensed corporations that using language models for investment recommendations, advice or research is high-risk, requiring extra safeguards and notification to the regulator. Singapore has no exact equivalent of that pre-notification duty. The privacy commissioner, the PCPD, published a model personal data protection framework for AI in June 2024, a checklist for employee use of generative AI in March 2025 and, in August 2026, Hong Kong’s first standalone guidance on agentic AI and privacy, which makes the deploying organisation responsible because an agent is not a legal person. And in October 2024 the Financial Services and Treasury Bureau issued a policy statement setting out a “dual-track” approach: promote AI adoption in finance while addressing its risks.
The engine of the Hong Kong model is the sandbox. The HKMA’s GenA.I. Sandbox, announced with Cyberport in August 2024, named its first cohort in December: ten banks, four technology partners and fifteen use cases, run on a dedicated environment at Cyberport’s AI Supercomputing Centre. Deputy Chief Executive Arthur Yuen called it “a risk-controlled testing ground for the industry”. The second cohort, announced in October 2025, grew to 27 use cases from 20 banks, with themes such as using AI to check AI and defending against deepfake fraud. When its results were presented in October 2026, more than 60 per cent of use cases involved AI agents, half of them multi-agent.
In March 2026 the model went cross-sector. The HKMA, the SFC, the Insurance Authority and the Mandatory Provident Fund Schemes Authority launched a joint GenA.I. Sandbox++ offering supervisory guidance, technical support and free GPU access. Its first cohort, named in August, had 36 use cases from 30 institutions and 27 technology partners, with an explicit focus on agentic AI: end-to-end onboarding, payments, insurance claims and AI that oversees other AI. No Singapore regulator offers free, regulator-brokered compute on that scale.
Then, in October and November 2026, Hong Kong moved ahead of Singapore on agent-specific banking guidance. On 15 October the HKMA announced its ART framework, for Accountability, Resilience and Trust, bundling updated supervisory guidance on responsible AI adoption, a training-data sharing platform with Cyberport and the University of Hong Kong, and continued agentic testing. At the Finance + A.I. Forum the next day, Executive Director Carmen Chu argued that “effective oversight need not mean human involvement in every step”, a frank acknowledgement that line-by-line human review cannot scale to agents.
The 2 November discussion paper is the most advanced supervisory thinking on agents in banking that I have read anywhere. Its organising idea is intent-anchored accountability: “An agent should operate only in pursuit of an approved business intent”, within a defined delegation of authority, with accountability resting on named individuals and undiluted by third-party models or tools. It proposes controls enforced technically through an “agent harness” rather than by prompts alone; an effective kill switch; one accountable owner per deployment; materiality factors that include reversibility; a central registry of material agents; and Know Your Agent for external agents, with their content treated as untrusted input. It is preliminary and non-binding, and first-round responses are due before 1 February 2027. But no Singapore document proposes a regulator-held agent registry, and MAS has deferred its agentic consultation to 2027.
Hong Kong’s weaknesses are real too. Its regulation is fragmented: the HKMA for banks, the SFC for licensed corporations, the Insurance Authority for insurers, each tailored, none unified, and the Insurance Authority has no standalone AI guideline that I could find. Its privacy law still lacks mandatory breach notification, which Singapore has had since 2021. It has almost no third-party assurance infrastructure and no tester accreditation. Its proposed copyright exception for text and data mining was shelved after pushback from rights holders, while Singapore’s has been in force since 2021.
Its strengths are just as real, and they are mostly about the state. Cyberport’s supercomputing centre provides around 3,000 petaflops of Hong Kong’s roughly 5,000, with a new data facility cluster planned to reach 180,000 petaflops by 2032. A HK$3 billion subsidy scheme pays a large share of compute costs for universities and firms. A HK$1 billion AI Research and Development Institute, chaired by Harry Shum, is mandated to help build governance standards. The September 2026 Policy Address created a seven-point AI risk-governance strategy and a new Commissioner for AI under the Digital Policy Office, with recruitment from the second quarter of 2027, and promised society-wide safety guidelines for AI agents in 2027. On political will and public money, Hong Kong is not behind.
And then there is the mainland. Beijing’s AI rules, from algorithm filing to the mandatory labelling of AI-generated content in force since September 2025, do not apply in Hong Kong under “one country, two systems”. But their shadow does. Hong Kong’s first home-grown large language model, HKGAI V1, launched in February 2025, was built on DeepSeek and is used across more than seventy government departments. Tencent Cloud sits on the HKMA’s Agentic Finance Expert Group. Meanwhile, by provider policy, major American models are restricted in Hong Kong, and Singapore has used access to both American and Chinese frontier models as a pitch for finance talent. When HSBC announced a new AI centre in Singapore in July 2026, the HKMA reportedly questioned the bank about it, despite Hong Kong being its largest market.
The fairest summary is the one a law firm offered at mid-year. Mayer Brown observed that Singapore builds assurance infrastructure while Hong Kong intensifies supervision under existing regimes, and that both converge on least privilege for agents, human oversight, logging and third-party risk. Hong Kong is a gateway; Singapore is a bridge. A bridge is valuable because neither bank of the river owns it.
Hong Kong is ahead on agents in banking and behind on assurance. Singapore is the reverse on the first and ahead on the second. Each city is the other’s missing half.
VII. Two Cities, Side by Side

The table below compresses the comparison. It reflects the position on 20 November 2026; several rows will move within months.
| Dimension | Singapore | Hong Kong |
|---|---|---|
| Philosophy | Test before you legislate; assurance as infrastructure; “pragmatic balance” | Dual-track: promote adoption, manage risk; sandbox, learn, then guide |
| Lead instruments | Model AI Governance Frameworks (2019, 2020, GenAI 2024, Agentic 2026); AI Verify; MAS AI Risk Management Guidelines | HKMA principles (2019) and ART framework (2026); SFC GenAI circular (2024); PCPD frameworks; DPO GenAI guideline |
| Testing model | Independent testers matched with real deployers across sectors and borders; open-source toolkits; accredited testers | Regulator-run sandboxes on state compute, finance only; little third-party assurance |
| Finance regulator stance | One cross-sector instrument for all institutions and all AI, phased to October 2027 and October 2028 | Split by regulator; HKMA guidance and SFC circulars, each tailored; SFC pre-notification for high-risk use |
| Agentic guidance | IMDA agentic framework (Jan 2026, updated May); SAFR runtime design (Jul 2026); MAS agentic consultation deferred to 2027 | HKMA discussion paper (2 Nov 2026): intent-anchored accountability, agent harness, kill switch, central agent registry; PCPD agentic privacy guidance |
| Know Your Agent | MAS to facilitate a KYA protocol under a new phase of SAFR, with an Agentic Money Sandbox (19 Nov 2026) | Risk-based KYA proposed in the discussion paper; Project SAGE workstream on agent identity (2 Nov 2026) |
| Legal backbone | No AI statute; sector regulators plus PDPA (breach notification), election deepfake ban, online safety law | No AI statute; sector regulators plus PDPO (no mandatory breach notification); Commissioner for AI from 2027 |
| Compute | Constrained; private hyperscalers; national supercomputing centre; new US chip licensing applies | State-built supercomputing centre, about 3,000 petaflops, planned 180,000 by 2032; HK$3bn subsidy |
| Geopolitical position | Neutral convener; access to US and Chinese models; hosts US-China safety dialogue | Gateway to the mainland; DeepSeek-based government model; limited US model access |
| Main weaknesses | No stop power outside regulated sectors; voluntary assurance; no general incident-reporting duty | Fragmented regulators; dated privacy law; thin assurance market; growing mainland dependence |
Read the table across, not down. Every Singapore weakness has a Hong Kong strength opposite it, and the reverse.
VIII. The Panel in Four Rooms
A story this flattering to one city needs cross-examination. As usual, the panel sits in four rooms: the public sector, the private sector and the labs, the academy, and the governance answer.
Room one: the public sector
Governments are the haven’s customers and its competitors. Josephine Teo has made the economic case most explicitly. Speaking at a Partnership on AI event in February 2026 on closing what she called the global assurance divide, she framed safety assurance as a competitive advantage rather than a cost, described the agentic framework as a live document, and said Singapore could not build the assurance ecosystem alone. A trust haven is only as valuable as the number of jurisdictions that accept its stamp.
Which is why the room is not entirely friendly. Washington’s position, voiced by Kratsios at Delhi, is that it rejects global governance of AI and offers American technology instead. Beijing has a rival offer. At the World AI Conference in Shanghai in July 2025, Premier Li Qiang proposed a World AI Cooperation Organization, and the Turing laureate Andrew Yao, warning that AI “lacks theoretical guarantees”, said he wanted Shanghai to become one of the world’s most important hubs for AI safety governance. Singapore is not the only city bidding to host the conversation. It is the only serious contender that sits inside neither superpower’s sphere, and that is both its advantage and its fragility.
A European government brief adds a sharper note. The research brief by Jasper Verstappen, published by the Netherlands Enterprise Agency in February 2026, credits Singapore’s adaptability, sector-tailored standards and use of industry expertise. It also names the two weaknesses I hear most often from European colleagues: limited enforceability compared with the EU AI Act, and the risk of industry capture, since those with most influence over voluntary rules often have the least interest in strict ones. A haven designed, tested and accredited with industry has to work hard to show it is not captured by industry.
The German Marshall Fund’s analysts put a ceiling on the whole idea. Middle powers, they wrote in October 2026, live in “managed interdependence”: they have leverage in minerals, chips and rule-setting, and they name Singapore as a leader in collective rule-setting, but they still depend on American and Chinese models, cloud and capital. A haven can set the terms of trust. It cannot set the terms of supply.
Room two: the private sector and the labs
The most striking voices for a trust haven come from inside the labs, and so do the most striking challenges. Jack Clark, Anthropic’s co-founder and head of policy, told NPR in September 2026 that slowing down is a collective-action problem, and named the missing piece bluntly: “We don’t have a third party that can validate what we’re saying”. That is a job description for the institutions Singapore has spent four years building. Mustafa Suleyman of Microsoft AI, rejecting the pursuit of superintelligence at any cost, wrote that “Accountability and oversight are to be welcomed when the stakes are this high”. Jensen Huang, the chip seller every government courts, said on the Dwarkesh Podcast in April 2026, as The Edge Singapore reported, that “having a research dialogue is probably the safest thing to do” between the United States and China, and called that dialogue glaringly missing. A neutral room is something industry increasingly wants, not merely tolerates.
But the sharpest challenge to Singapore’s voluntarism also comes from a lab. In June 2026 Dario Amodei published “Policy on the AI Exponential”, arguing that binding US federal regulation is no longer optional and proposing an aviation-style regime with mandatory third-party testing above a compute threshold, and the power to block or reverse deployment. Frontier AI, he wrote, “cannot safely be fully entrusted to either governments or companies”. This cuts both ways. It vindicates Singapore’s bet that independent testing is the core institution of AI governance. It also says, from the heart of the industry, that voluntary testing is not enough. A haven whose stamp anyone can ignore is a library, not a gate.
The evaluators raise a deeper problem. Alex Meinke of Apollo Research argued in July 2026 that “Final-checkpoint evaluations will be insufficient to assess scheming risks.” Deceptive behaviour can emerge during training, and independent assessors need access to intermediate checkpoints that developers have little commercial reason to share. Singapore tests applications built on frontier models, not the training runs that produce them. That is a deliberate and defensible choice for a small state, but it means the haven can vouch for what a bank’s chatbot does, not for what the model underneath it might be capable of. Rumman Chowdhury, whose Humane Intelligence partnered with IMDA on red-teaming, has long warned that “companies write their own tests and they grade themselves”,, and reporting in San Francisco in October 2026 found that frontier labs largely choose their own evaluators from a small, interconnected pool. Accrediting testers against public methods is a partial answer. It is not yet a full one.
Room three: the academy
Singapore’s own scholars are its most rigorous critics, which is itself a sign of a healthy system. Aditya Vikram Kashyap’s ISEAS Perspective of September 2026, titled “Who Can Actually Stop the Machine?”, makes the enforceability argument better than anyone. Singapore has “no omnibus statute and no cross-sectoral AI enforcement agency”. The binding layer is the ordinary authority of sector regulators, which is real in finance, where MAS licenses and inspects, and largely absent elsewhere. If a firm that no sector regulator supervises deploys a harmful system in hiring or scoring, harms are handled after the fact, and only if victims discover AI was involved. Kashyap accepts that choosing assurance before authority is defensible. He insists that it is a choice not to create stop power, at least for now, and he wants an identifiable body able to say no for the highest-consequence uses.
Manoj Harjani of RSIS asked in 2025 where Singapore’s AI regulation is headed, doubting that a light-touch model can handle deliberate misuse and adversarial attack, and noting that frontier-risk and superintelligence debates are largely sidestepped in official discourse. Jason Grant Allen of SMU responded that light touch is a political-economy choice, not a neutral one: soft rules attract capital and shift power and accountability towards firms, and the story of unregulated innovation ignores how much the state steers. With Jane Loo and Jose Luna, Allen has also questioned whether Singapore’s model “translates into economies and societies of a different scale”. A method that works in a compact city-state with a trusted civil service may not survive export to a federation of hundreds of millions.
Lim How Khang, also of SMU, offers the most constructive framing. In “Assurance as Infrastructure” he describes three layers, an enforced baseline of data, cyber and online-harms law, targeted statutes, and a voluntary assurance layer on top, and calls the result sovereignty-by-assurance. He is clear about its limits. A deployer-centred model struggles with failures in foundation models and offshore attacks that have no local deployer. The voluntary layer might harden through procurement, insurance and due diligence, but that is unproven. And, most importantly for a trust haven, assurance is worthless unless other jurisdictions recognise it.
Then there is the challenge to the frame of this essay itself. Arvind Narayanan and Sayash Kapoor, authors of “AI as Normal Technology”, argue that extraordinary, liberty-restricting measures against AI are costly and of limited use because AI has no physical chokepoint, and that “relying on nonproliferation as the primary defense creates fragility”. Build resilience instead, through normal, polycentric policy. Their view supports Singapore’s incremental method, and questions whether “superintelligence arms race” is the right lens at all. I think it is, because the builders say so and fund accordingly. But they are right that a haven built only for apocalypse would neglect the ordinary harms that arrive first.
Finally, Stuart Russell. On a Singapore stage he argued that developers should carry the burden of proving their systems safe, as in aviation, medicine and nuclear power: “Don’t wait for Chernobyl”. On the same panel Zhang Yaqin of Tsinghua called for AI agents to be registered and AI content labelled, a Chinese scientist foreshadowing Know Your Agent at a Singapore summit. And in Paris in February 2026, Russell observed that middle powers beyond the United States and China were becoming more open to stricter regulation. If the middle powers are moving towards binding rules, a voluntary-first haven must ask whether it is leading them or being overtaken.
Room four: the governance answer
Put the rooms together and a fair verdict emerges. Singapore has the best method in the world for producing evidence about AI systems: open tools, independent testers, published runtime designs, international standards, and a willingness to publish its own failures in Thai. What it lacks is the authority to make that evidence decisive. In finance, MAS supplies the authority, and the guidelines are, as Kashyap notes, effectively binding on the firms it supervises. Outside finance, the incentive to be tested is mostly reputational.
Hong Kong has more of the opposite: supervisory instinct, a willingness to float a registry and a kill switch, state money and state compute, but far less of the independent assurance machinery that would make its supervision verifiable by outsiders. The honest conclusion is that neither city has yet built a trust haven. Singapore has built most of one. The rest is not more frameworks. It is teeth, recognition and interoperability.
Method without authority is advice. Authority without method is theatre. A trust haven needs both.
IX. What a Trust Haven Still Needs
Every regime I examine faces the same question, which I call the enforceability test: when a deployment is about to cause serious harm, can anyone stop it, and can anyone outside the firm verify that it was stopped? Singapore passes the test in finance, where MAS can require an institution to limit, suspend or replace an AI service. It does not yet pass it across the economy. The fix need not be an omnibus AI Act. It could be a narrow statutory backstop, an identifiable authority with power to halt the highest-consequence uses, as Kashyap proposes, paired with a general duty to report serious AI incidents, which Singapore still lacks. A haven that cannot say no is a showroom.
Second, mutual recognition. Lim How Khang is right that assurance is worthless unless others accept it. ISO/IEC 42119-8 is the right instinct: export the method as a standard. The next step is bilateral and regional recognition of accredited test results, so that an application tested in Singapore under a published method does not need to be retested from scratch in London, Tokyo or Brussels.
Third, runtime evidence. Testing before deployment is necessary and insufficient, because agents change their behaviour as their tools, data and instructions change. The Agentic Testing Starter Kit already asks testers to monitor and re-test. SAFR already sketches a gate and a tamper-evident log. What is missing is a requirement that the high-consequence actions of agents leave a record the agent cannot edit, in a form a supervisor or an accredited tester can read. In this week’s companion piece on governance platforms I argued that most commercial tools describe and score risk but cannot refuse an action. A trust haven must make refusal, and the evidence of refusal, the norm. That is the layer TrustOS, the agentic governance architecture I designed, was built to occupy: a disposition before every consequential action, and a receipt after it. Whose design fills that layer matters far less than that it exists and can be audited.
Fourth, agent identity standards, and here the two cities must stop running in parallel. MAS’s know-your-agent protocol and the HKMA’s registry and KYA approach address the same problem from two sides: MAS from the protocol an agent presents at the door, Hong Kong from the record a regulator keeps of the agents inside. They are complements, not competitors. An agent that presents a Singapore credential to a Hong Kong bank, or the reverse, will be routine within a few years. If the fields do not map, both cities will have built walled gardens in a region that needs open roads. The Singapore Consensus put agent identity verification on the research agenda in 2025. The two monetary authorities should now agree a common minimum: identity, principal, authority, mandate, provenance, and a reference to a registry entry where one exists.
Fifth, and this is the hardest, the haven must engage with the frontier, not only with its applications. Apollo’s point stands. A small state cannot evaluate training runs on its own. But it can host the evaluators, set the accreditation standard they work to, and offer itself, as it did with the Consensus, as the room where American and Chinese labs agree what should be tested and how. The Frontier AI Resilience and Security Taskforce, due to report in 2027, is an opening. It should be used.
Behind all five sits a larger question, the one I keep coming back to in this series and in Genesis before it. We are approaching a fork. Down one road lies the Star Trek future: abundance, agents that extend human capability, institutions that keep powerful systems answerable. Down the other lies Mad Max: concentration, capture, systems that act faster than anyone can hold them to account. Markets will not choose the road; they will price whichever road we are on. I believe that only governments can set the trajectory for the Fork. That is why a small government’s choices matter far beyond its size. Singapore cannot decide which road the superpowers take. It can build, in public, a working model of the road worth taking, and make it cheap for others to follow.
Yuval Noah Harari, speaking in Beijing in 2025, identified the paradox at the centre of the race: the builders say they cannot trust rival humans, yet they expect to trust the superintelligence they build. His answer was that “we need to trust other humans more than we trust AI”. A trust haven is, in the end, an institution for exactly that: a place where humans who do not trust each other can agree on evidence about machines.
The haven needs a backstop, recognition, runtime receipts, shared agent identity and a seat at the frontier. Above all it needs governments willing to steer.
X. What to Do on Monday Morning
For boards in Singapore. Read the MAS guidelines as binding even where they say “should”, because your supervisor will. Build the AI inventory and materiality assessment now; October 2027 is closer than your next budget cycle. For every agent that can move money or change customer records, ask three questions: who is its accountable owner, what is its mandate in writing, and who can stop it, how fast. If you buy agentic services from third parties, check that your contract lets you suspend them.
For boards in Hong Kong. Respond to the HKMA discussion paper before 1 February 2027, and do it with data, not lobbying. Start keeping the registry the paper describes, internally, whether or not the central version arrives: purpose, owner, access, dependencies, stop mechanism, change history. If you operate in both cities, design one agent record that satisfies both, because you will not want to maintain two.
For regulators in both cities. MAS and IMDA: decide where the stop power lives outside finance, and introduce a serious-incident reporting duty before an incident forces one. Bring the agentic consultation forward if adoption outpaces the 2027 timetable. The HKMA: pair the registry with an assurance regime, or the registry becomes a list of claims nobody can test. Both: open a working channel on know-your-agent fields now, while both designs are still soft.
For builders. Put your application through the Global AI Assurance Sandbox or an accredited tester, and publish the summary. Test in every language your users speak; the Thai case study is the template. Build the gate before the action and the receipt after it into your architecture, not as an afterthought. Design your agents to present identity, principal, authority and mandate on request, because both cities are about to ask.
For everyone. Follow Chia’s advice and do not hand your personal agent the passwords to your bank account. Not yet.
Governance happens on Monday morning, in inventories, contracts and logs, not in keynotes.
Coda: The Season of Light
Dickens opened his tale of two cities with a run of contradictions: wisdom and foolishness, belief and incredulity, the season of Light and the season of Darkness. The age of agents has the same split personality. The same week brought tokenised settlement, agentic money, a know-your-agent protocol, and a central banker warning the public not to trust any of it with their passwords yet.
Singapore’s achievement is real. It turned the weakness of a small state into a strategy: if you cannot win the race, become the place where the racers’ machines are checked, and make the checking open, independent and good enough that both sides accept it. Hong Kong’s mirror shows what Singapore still lacks: the willingness to keep a register, to name a kill switch, to say no. Neither city has finished. Both are, for once, asking the right question at the same time.
The best of times, the worst of times. The difference, as ever, will be made by those who choose to govern.
Library
Singapore: frameworks, strategy and law
- Allen and Gledhill, Singapore releases Asia’s first Model AI Governance Framework (Feb 2019)
- PDPC and IMDA, Model AI Governance Framework, second edition (Jan 2020)
- MDDI, National Artificial Intelligence Strategy unveiled (Nov 2019)
- MDDI, Launch of National AI Strategy 2.0 (Dec 2023)
- Morgan Lewis, Singapore expands its AI governance approach to include generative AI (Jul 2024)
- MDDI, Singapore launches Model AI Governance Framework for Agentic AI (Jan 2026)
- Allen and Gledhill, IMDA updates Model AI Governance Framework for Agentic AI (Jul 2026)
- CSA, Addendum on securing agentic AI systems (Oct 2025)
- Allen and Gledhill, PDPC advisory guidelines on personal data in generative AI (Jul 2026)
- Malay Mail, Singapore passes law banning deepfakes of candidates (Oct 2024)
- IMDA, ASEAN Working Group on AI Governance
- MDDI, Singapore-UK agreement on AI safety (Nov 2024)
- Enterprise Singapore, Singapore champions new global AI testing standardisation efforts (Apr 2026)
- Prime Minister’s Office, DPM Gan Kim Yong at the Singapore FinTech Festival 2026 (18 Nov 2026)
Singapore: testing and assurance
- MDDI, Singapore launches first AI testing framework and toolkit (May 2022)
- AI Verify Foundation, What is AI Verify
- AAP/PR Newswire, AI Verify Foundation launched (Jun 2023)
- Allen and Gledhill, Generative AI Evaluation Sandbox for Trusted AI (Oct 2023)
- MDDI, Minister Josephine Teo at ATxAI: Project Moonshot (May 2024)
- NTU, Digital Trust Centre designated as Singapore’s AI Safety Institute (May 2024)
- MDDI, Singapore announces new AI safety initiatives (Feb 2025)
- IMDA via PR Newswire, Insights from technical testing of real-world GenAI applications (May 2025)
- AI Verify Foundation, Global AI Assurance Pilot main report (May 2025)
- AI Verify Foundation, Global AI Assurance Sandbox
- Techgoondu, Singapore expands AI assurance pilot to test AI agents and prompt injections (Jul 2025)
- AI Verify Foundation, Sandbox case study: KBTG and Vulcan
- AI Verify Foundation, AI Tester Accreditation Programme
- Allen and Gledhill, IMDA introduces new AI testing and cybersecurity initiatives (Oct 2026)
- IMDA, Starter Kit for Testing LLM-Based Applications for Safety and Reliability
- The Singapore Consensus on Global AI Safety Research Priorities (arXiv 2506.20702)
- FAR.AI, The Singapore Consensus
MAS and finance
- MAS, FEAT principles (Nov 2018)
- MAS, Veritas
- MAS, Project MindForge
- Allen and Gledhill, MAS information paper on AI model risk management (Dec 2024)
- MAS, Consultation on Guidelines on AI Risk Management (Nov 2025)
- MAS, Supervisory expectations on responsible AI adoption (7 Oct 2026)
- MAS, Safeguards for AI agents in finance (SAFR, 3 Jul 2026)
- MAS, SAFR white paper (PDF)
- MAS, Future of Finance Institute commences operations (3 Nov 2026)
- MAS, Keynote address by the Managing Director at the Singapore FinTech Festival 2026 (19 Nov 2026)
- The Straits Times via AsiaOne, Don’t give AI agents your financial accounts login details (19 Nov 2026)
- Fintech Singapore, AI agent payments trials (Nov 2026)
- Blakes, OSFI final Guideline E-23 on model risk management and AI (Sep 2025)
Hong Kong
- HKMA, High-level Principles on Artificial Intelligence (Nov 2019)
- Gibson Dunn, HKMA guidance on GenAI in customer-facing applications and AI in AML monitoring (Sep 2024)
- HKMA, GenA.I. Sandbox first cohort (Dec 2024)
- HKMA, GenA.I. Sandbox second cohort (Oct 2025)
- HKMA, GenA.I. Sandbox++ launch (Mar 2026)
- HKMA, GenA.I. Sandbox++ first cohort (Aug 2026)
- HKMA, ART framework (Oct 2026)
- HKMA, Carmen Chu at the Finance + A.I. Forum (Oct 2026)
- HKMA, Discussion Paper on Agentic A.I. Governance in the Banking Sector (Nov 2026)
- HKMA, Project SAGE and the Agentic Finance Expert Group (Nov 2026)
- Clifford Chance, SFC circular on generative AI (Nov 2024)
- PCPD, Checklist on Guidelines for the Use of Generative AI by Employees (Mar 2025)
- Reed Smith, PCPD guidance on agentic AI (Sep 2026)
- FSTB, Policy Statement on Responsible Application of AI in the Financial Market (Oct 2024)
- Hong Kong SAR Government, Policy Address 2026 (Sep 2026)
- Legislative Council, Background brief on AI development and governance (Jun 2026)
- The Standard, HKGAI V1 built on DeepSeek (Feb 2025)
- South China Morning Post, HKMA questions HSBC over Singapore AI centre (Oct 2026)
- Mayer Brown, AI Regulation in Singapore and Hong Kong: A Mid-Year Checkpoint (Jul 2026)
The arms race, the labs and the multilateral setting
- CSET, Helen Toner testifies before the Senate Judiciary Committee (Apr 2026)
- Brookings, Takeaways from the India AI Impact Summit (Feb 2026)
- WIRED via AITopics, Singapore’s vision for AI safety bridges the US-China divide (May 2025)
- MDDI, Minister Josephine Teo at the Singapore Conference on AI (Apr 2025)
- MDDI, Minister Josephine Teo on closing the global assurance divide (Feb 2026)
- Fortune, Dario Amodei’s “The Adolescence of Technology” (Jan 2026)
- ThePrint, Amodei calls for binding AI regulation (Jun 2026)
- NPR via KJZZ, Jack Clark on slowing AI as a collective-action problem (Sep 2026)
- Mustafa Suleyman, Towards Humanist Superintelligence (Nov 2025)
- Demis Hassabis on China’s AI gap (Jan 2026, secondary)
- Yicai Global, Andrew Yao on global AI safety governance (Jul 2025)
- Sixth Tone, China proposes a global AI governance body (Jul 2025)
- German Marshall Fund, Middle powers’ bounded ambition in AI (Oct 2026)
- The Star/AFP, Stuart Russell on middle powers and stricter rules (Feb 2026)
- Computer Weekly, AI safety cannot wait for a Chernobyl moment (ATxSummit)
- Yuval Noah Harari, The paradox of trust (Beijing, Mar 2025)
Assurance, evaluation and the critics
- Aditya Vikram Kashyap, Who Can Actually Stop the Machine? ISEAS Perspective 2026/70 (Sep 2026)
- Manoj Harjani, Where Is Singapore’s AI Regulation Headed? Issues in Science and Technology (2025)
- Jason Grant Allen, A Difficult Act to Maintain, Issues in Science and Technology (Nov 2025)
- Allen, Loo and Luna, Governing Intelligence, Cambridge Forum on AI: Law and Governance (2025)
- Lim How Khang, Assurance as Infrastructure, Tech For Good Institute (Aug 2026)
- Jasper Verstappen, AI Governance in Singapore, ESF Research Brief No. 3 (Netherlands Enterprise Agency, Feb 2026)
- CNBC, Singapore is not looking to regulate AI just yet (Jun 2023)
- Alex Meinke (Apollo Research), We need third-party training-run evaluations (Jul 2026)
- Stanford HAI, Strengthening AI accountability through better third-party evaluations (Nov 2024)
- San Francisco Standard, Who is watching the AI companies? (Oct 2026)
- Arvind Narayanan and Sayash Kapoor, Do AI risks require extraordinary government intervention? (2026)
Companion pieces on Genesis
- Singapore’s Model AI Governance Framework for Agentic AI (Feb 2026)
- Safeguards for Agentic Finance at Runtime (SAFR) (Jul 2026)
- MAS Finalises AI Risk Guidelines: What Actually Changed (Oct 2026)
- Control not Pause (Oct 2026)
Dr Luke Soon has spent three decades in AI and now focuses on AI governance and safety. He is the author of Genesis: Human Experience in the Age of Artificial Intelligence and Synthesis: SuperIntelligence Protocol, and believes that only the governments of the world can set the trajectory for the Fork.
Short-term turbulence for long-term abundance.


Leave a Reply