Six Regimes, One Agent: The November Update

On 7 October I compared six AI regimes for financial institutions, the day Singapore’s Guidelines on AI Risk Management were issued. Six weeks later the map has widened, and one of the six has changed its mind. This update widens the lens to 89 jurisdictions and international instruments and 27 frameworks, corrects the record where the record has moved, and asks again the question that matters most for banks and insurers: who is actually governing the agents?

An update to Six Regimes, One Agent (7 October 2026). Interests declared: I co-authored IMDA Singapore’s Model AI Governance Framework for Agentic AI, and I have a commercial interest in TrustOS, a runtime governance architecture. The wider comparison draws on a research briefing compiled on 15 November; the key facts below were re-checked on 15 and 16 November, with research and drafting help from Claude, a model made by Anthropic. Facts are stamped as of 16 November 2026.

The Two-Minute Version

  • One of my original six has moved. On 10 November Australia accepted all 18 recommendations of its Medicare review, including a new AI safety regulator and mandatory, time-bound incident reporting. In October I filed Australia under “existing rules, clarified by letters”. That is no longer the whole story.
  • Asia did more in a month than most regions did in a year. The Philippines adopted a national AI governance framework by executive order on 12 November. Malaysia says its AI Governance Bill is finished and heads to Parliament early next year. China’s top planner promised on 3 November to speed up a comprehensive AI law.
  • One correction to the wider record: Vietnam’s list of high-risk AI systems is out. Decision 33 names 46 systems in six sectors, banking among them, with compliance due by 1 March 2027.
  • Binding law is winning as a template and losing as a timetable. Almost everyone borrows the EU’s risk tiers; almost no one enforces them at EU scale or on EU time, and the EU itself has pushed its high-risk duties to December 2027 and August 2028.
  • The agent test has two new entrants. In October only Singapore wrote agent-specific expectations. China has since issued guidance on AI agents, and Australia’s reforms will require reporting of misaligned behaviour. Singapore is still furthest ahead.
  • For multi-jurisdictional institutions, my October advice mostly stands: treat Singapore’s guidelines as the agent baseline. But Vietnam’s banking deadline and Australia’s coming incident clock now ask for things the Singapore guidelines do not.

I. Why an Update, and What Changed in the Method

Six Regimes, One Agent was deliberately narrow. It compared supervisory expectations for financial institutions in Singapore, the EU, the UK, the US, Hong Kong and Australia, through six questions and one hard case: the AI agent. It said at the time that several regimes were in motion and the position could change quickly. It has.

This update does two things. First, it records what has changed in the original six, and in the region around them, since 7 October. Second, it widens the frame. A research briefing compiled on 15 November compared 89 jurisdictions and international instruments, and 27 frameworks and standards, across the same dimensions: legal force, scope, obligations, oversight and penalties. I have not reproduced its tables here. I have used it as a map, re-checked the entries that matter most for an Asia-based institution, and corrected two that were out of date within hours of being written.

A word on confidence. The briefing’s authors were careful to say that most rows rest on reputable secondary sources rather than each regulator’s own site, and they flagged their conflicts rather than smoothing them over. I have kept that discipline. Where I could confirm a fact this week, I say so. Where I could not, it is in Section X.

II. Six Weeks, Dated

Here is what moved between the first comparison and this one, in order.

DateJurisdictionWhat happened
7 OctSingaporeMAS issues its Guidelines on AI Risk Management, the subject of the original comparison.
8 OctIndiaThe IT minister says MeitY will publish a regulation-focused AI consultation paper within a month, a shift away from relying only on existing laws and voluntary compliance.
13 OctSouth KoreaCabinet approves a further amendment to the AI Basic Act on training data, as reported by Korean press and MLex. The fine grace period still has no official end date.
29 OctMalaysiaThe digital minister says the AI Governance Bill is complete, will go to Cabinet, and should be tabled in Parliament by early 2027.
3 NovChinaThe National Development and Reform Commission says it will speed up a basic, comprehensive AI law, with safety standards across the full life cycle.
10 NovAustraliaThe government accepts all 18 recommendations of the Medicare rapid review: an AI safety regulator, mandatory time-bound reporting, independent evaluation of frontier models and possible criminal-law reform.
12 NovPhilippinesExecutive Order 131 adopts a national AI governance framework and creates a National AI Council chaired by the Executive Secretary. No penalties; sector regulators stay in charge.

Two things stand out. The movement is almost all in Asia-Pacific. And none of it is in the direction of less governance. Even the lightest-touch step on the list, the Philippine order, adds an institution and a reporting cycle where there was none.

III. Correcting the Record

Caricature of an officer correcting a large ledger with a red pencil while a robot holds an amber lamp over the page

Two entries need correcting, one in my own October piece and one in the wider briefing.

Australia. In October I described Australia as applying existing rules, clarified by APRA’s letter to industry of 30 April. That was accurate for the prudential layer, and it still is: APRA’s expectations stand. But the government’s acceptance of the Medicare review on 10 November changes the economy-wide picture. Australia now intends to have a dedicated AI safety regulator, mandatory and time-bound reporting of AI incidents and misaligned behaviour, independent evaluation of frontier models, and work on criminal penalties for serious misconduct. None of this is law yet, and the detail will matter. But the direction has reversed. A country that had shelved mandatory guardrails in December 2025 is now building the most incident-focused regime in this comparison. The briefing, compiled on 15 November, still listed Australia as voluntary with its guardrails dropped. It was right about the guardrails and wrong about the direction.

Vietnam. The briefing said the prime minister’s list of high-risk AI systems was still unpublished. It was published in July. Decision 33/2026, issued on 30 June and in force from 15 August, names 46 high-risk systems across six sectors: education, ethnicity and religion, healthcare, banking, transport and judicial proceedings. Listed systems need either a self-assessment or third-party certification, depending on the system, and compliance is due by 1 March 2027. Many systems on the list are high-risk only where an AI output is the sole basis for a decision, without meaningful human review. For banks, that turns Vietnam from a watch item into a deadline.

Neither correction is a criticism of the people who did the work. Both show the same thing: in this field a snapshot ages in days, and a comparison that is not re-checked before publication will mislead.

IV. Asia-Pacific Is Now Five Postures, Not One

The original six included three Asia-Pacific regimes: Singapore, Hong Kong and Australia. Widen the lens and the region sorts into five postures.

PostureWhoWhat it means for an institution
Binding, risk-tiered lawSouth Korea (in force January 2026, fines deferred); Vietnam (high-risk list out, compliance by 1 March 2027)Hard floors with dates. Korea’s fines are capped at KRW 30 million and suspended for at least a year; Vietnam’s conformity duties for listed systems are the sharper obligation.
Measure-by-measure regulation, law comingChinaFiling, labelling and sector rules now, including guidance on AI agents issued in May; a comprehensive law promised but not yet drafted in public.
Moving towards a regulatorAustraliaAPRA’s prudential expectations today; an AI safety regulator and incident clock to come.
Supervisory and voluntary, with runtime toolsSingapore, Hong KongNo AI statute. In Singapore, MAS expectations are functionally binding for regulated firms, backed by the agentic framework and the SAFR runtime model.
Coordination or promotion first; drafts in motionPhilippines, Japan, India, Indonesia; Thailand and Malaysia draftingInstitutions and principles now, duties later. Thailand’s draft is the most EU-shaped in ASEAN; Malaysia’s bill would create a Central AI Authority.

For a group operating across the region, the consequence is a compliance stack, not a rulebook. Vietnam and Korea set the hard legal floors. MAS sets the most demanding supervisory expectations. Australia will set the incident clock. Thailand and Malaysia are 2027 risks. The cheapest way to satisfy all of them at once is the same as it was in October: evidence produced by the running system, not documents produced for the regulator.

The Philippine order deserves a word of its own, because it is easy to dismiss. It imposes nothing on the private sector and could be revoked by the next president. But it is unusually explicit about protecting workers and learners through the AI transition, and it creates a quarterly reporting loop to the President. On the human side of governance, the side I call HX, it asks a question most statutes skip.

V. The Frameworks Layer Is Where Institutions Actually Live

Caricature of a robot balancing on a tall wobbly stack of binders in which only one binder carries a brass seal

The briefing’s second table compares 27 frameworks and standards, and for practitioners it is the more useful one. Across 89 jurisdictions the statutes differ wildly. The frameworks companies actually build to are few and portable: the NIST AI Risk Management Framework as the risk process, ISO/IEC 42001 as the management system, IMDA’s Model Governance Framework family and AI Verify as the testing layer, and in finance, MAS’s FEAT principles and now the AI Risk Management Guidelines.

One finding deserves to be better known. Of everything compared, ISO/IEC 42001 is the only AI management-system standard that a third party can certify. NIST’s framework has no certification by design. AI Verify says plainly that it does not certify and does not guarantee a system is risk-free. And, per the briefing, 42001 was not yet cited as a harmonised standard under the EU AI Act as of September, so a certificate does not buy presumption of conformity in Europe.

That matters for the assurance question I keep returning to. A board that asks “are we certified?” will get a yes or no about a management system. It will not get an answer about whether a specific agent did what it was allowed to do yesterday. Certification is a floor for process. It is not evidence about behaviour.

VI. The Agent Test, Revisited

In October the agent test sorted the six regimes into three groups, and Singapore stood alone in writing agent-specific expectations: inventory, monitoring of reasoning, actions and tool calls, and the ability to stop an agent. That remains the most complete set in this comparison.

Two entrants have joined since, from very different directions. China’s Cyberspace Administration, NDRC and MIIT jointly issued implementation guidance on AI agents in May, which the wider briefing did not pick up; Chinese rule-making is now addressing agents as such, not just generated content. And Australia’s accepted reforms will require reporting of misaligned behaviour, which is, in effect, an obligation to notice what agents do and say so. California’s frontier-model law already carries incident duties, and Korea’s Act places lifecycle duties on the largest systems.

Everywhere else, agents are governed awkwardly or not at all. The EU AI Act reaches them through annexes written before the agentic turn. Most of the 89 rows have no agent-specific provision. And, as the briefing notes, not one regime imposes a legal duty to reverse an AI decision. The closest analogues are all after the fact: incident reports, suspension, withdrawal.

So the test I proposed in October still sorts the field, and I would now add one question to it. Not only “can you see and stop the agent?”, but “who outside your organisation will hear about it, and how fast?” Australia’s answer is about to become “a regulator, on a clock”. Singapore’s is “your supervisor, when it asks”. Most of the world’s is “nobody, unless it reaches the papers”.

VII. What Changes for a Multi-Jurisdictional Institution

Caricature of a robot strolling across painted-on checkpoint lines towards one real brass gate with a traffic signal

In October I recommended treating the MAS guidelines as the baseline for agents everywhere, because no other regime asked for less and none yet asked for more. For agent inventory, monitoring and the ability to stop, that still holds. But two regimes now ask for something the Singapore guidelines do not.

  • Vietnam asks for conformity assessment by a date. If a group runs a listed banking system in Vietnam, Decision 33 requires a self-assessment or certification before 1 March 2027. That is a pre-deployment legal duty, not a supervisory expectation, and it has a deadline.
  • Australia will ask for incident reporting on a clock. Once legislated, mandatory, time-bound reporting of AI incidents and misaligned behaviour will be stricter on timing than anything in the Singapore guidelines. The groups that can meet it will be the ones whose agents already produce runtime evidence.

The practical revision to my October table is therefore small but real. Agent inventory and monitoring: Singapore still sets the highest bar. Pre-deployment conformity for listed systems: Vietnam, from March. Incident clocks: Australia, once legislated, with the EU’s clocks waiting behind the deferred high-risk duties. Everything else in the October table stands.

VIII. The Usual Panel, in One Table

VoicePositionWhat they would say about this update
Monetary Authority of SingaporeAgent inventory, monitoring and the ability to stop, applied proportionately to all regulated firmsWould note that Singapore wrote agent expectations first and that others are now following
Anthony Albanese, Prime Minister of AustraliaAccepted all 18 recommendations of the Medicare reviewWould say Australia learned from an incident and moved from voluntary guidance to a regulator
Gobind Singh Deo, Malaysia’s Digital MinisterA horizontal AI Governance Bill with a Central AI Authority, to be tabled early 2027Would argue Malaysia is building enforcement capacity before the law lands, not after
Ashwini Vaishnaw, India’s IT MinisterA regulation-focused consultation paper as the base for future rulesWould say India is moving from voluntary compliance towards regulation, at its own pace
China’s National Development and Reform CommissionSpeed up a comprehensive AI law with full life-cycle safety standardsWould frame AI safety as a global problem needing international cooperation
Korea’s Ministry of Science and ICTLaw in force, fines deferred for at least a yearWould defend the grace period as adjustment time; critics at IAPP have called the design all roar and no bite
A chief risk officer of a regional bank (composite view, not a named person)One control set across many regimesWould ask for a single evidence stack that satisfies Hanoi, Singapore and Sydney at once

IX. What Would Change My Mind

  • If Australia’s legislation, when drafted, turned the accepted recommendations into voluntary guidance, my claim that it has reversed direction would be wrong.
  • If Vietnam extended its 1 March 2027 deadline or narrowed the banking entries, the pressure on regional banks would ease.
  • If ISO/IEC 42001 were cited as a harmonised standard under the EU AI Act, certification would carry more legal weight than I give it here.
  • If a second regulator issued agent-specific expectations as detailed as Singapore’s, I would stop describing Singapore as alone at the front.

X. What I Could Not Verify

  • The 89-row and 27-row tables are the research briefing’s, built mostly from secondary sources read on 15 November. I re-checked the Asia-Pacific entries that matter for financial institutions, not every row.
  • The status of the Council of Europe’s Framework Convention on AI: the briefing says it was not in force as of late September 2026; one secondary source claims it entered into force earlier. I could not settle this from primary sources and have left it out of the analysis.
  • Whether India’s promised consultation paper has been published. It was announced on 8 October for publication within a month; I found no confirmation of release.
  • The full text of China’s May guidance on AI agents, and whether it imposes duties or only encourages them.
  • Which specific banking systems appear on Vietnam’s list, and which require third-party certification rather than self-assessment.
  • The details of Australia’s new regulator, its reporting clock and its penalties. These await legislation.
  • The Philippine order’s publication date in the Official Gazette, on which press reports conflict.

XI. What to Demand on Monday Morning

Boards of regional banks and insurers

  • Ask whether any system the group runs in Vietnam appears on the Decision 33 list, and what the plan is for 1 March 2027.
  • Ask how fast the group could tell an Australian regulator about an agent incident today, and who would make the call.

Chief risk and compliance officers

  • Keep the MAS guidelines as the agent baseline, and add two overlays: Vietnam’s conformity assessment and Australia’s incident clock.
  • Do not treat an ISO/IEC 42001 certificate as evidence about agent behaviour. Ask for runtime evidence instead.

Regulators, starting in Singapore and ASEAN

  • Use the regional momentum. With Malaysia, Thailand and the Philippines all building institutions now, a shared ASEAN template for agent incident reporting would cost little and save every regional bank a duplicate effort.

Coda: The Map Gets Bigger, the Question Stays the Same

Six regimes became 89 rows in six weeks, and the instinct is to read that as fragmentation. Look closer and it is convergence of a particular kind. Almost everyone now agrees that someone must know where the AI is, who owns it and what happens when it fails. They disagree on who that someone is and how fast they must hear.

That is the question the agent forces. Singapore answered it first, through supervision. Australia is about to answer it through a regulator and a clock. Most of the map has not answered it yet. The next update to this comparison will be worth writing when the answer is not “nobody, unless it reaches the papers”.

Library

The original comparison and its sources

Australia

Philippines

Vietnam

Malaysia, Thailand, India, Korea, China

European Union and the United States


Discover more from Genesis: Human Experience in the Age of Artificial Intelligence | Synthesis: The Superintelligence Protocol

Subscribe to get the latest posts sent to your email.

One response

  1. Six Regimes, One Agent: How Singapore’s AIRG Compares with the EU, UK, US, Hong Kong and Australia – Genesis: Human Experience in the Age of Artificial Intelligence | Synthesis: The Superintelligence Protocol Avatar

    […] Update, 16 November 2026: Australia, Vietnam and the wider region have moved since this was written. See Six Regimes, One Agent: The November Update. […]

    Like

Leave a Reply

Short-term Turbulence for Long-term Abundance.