Hansard Explorer

ADJOURNMENT MOTION

The Leader of the House (Ms Indranee Rajah): Mr Speaker, Sir, I beg to move, "That Parliament do now adjourn."

Question proposed.

National Digital Identity Infrastructure for Artificial Intelligence Agents

Mr Speaker: Ms Tin Pei Ling.

8.36 pm

Ms Tin Pei Ling (Marine Parade-Braddell Heights): Mr Speaker, Sir, before I begin, I declare that I am working in a Monetary Authority of Singapore (MAS)-licensed Major Payment Institution that develops and delivers agentic finance solutions.

Sir, artificial intelligence (AI) is developing at an unprecedented pace. This year, agentic AI sits at the centre of many global discussions. Today, most use cases still involve a human instructing an agent, that is human-to-agent. But we should expect agent-to-agent interactions to scale soon enough, where one agent negotiates, transacts or collaborates with another without a human in between.

That autonomy demands new ways of governance. Our existing policies and tools were built for a world where humans decide and machines execute, and they will need rethinking.

Imagine a scenario in which an AI agent acts on your behalf, say, booking your travel, retrieving your medical records, negotiating a contract, or moving your money, how will the counterparty know who exactly is that agent? Who is really executing the transaction? And who answers for it if something goes wrong? No one can say with certainty in most cases today.

Contrast this with how we treat human users online. For example, when we log in to a Government service, Singpass tells the agency end who we are, thereby allowing the agency to confidently process the request.

To this end, I believe there is scope for a National Digital Identity Infrastructure for AI agents. To get ready for the agentic future.

AI safety and governance have been at the centre of many discussions in recent months. There is good reason. Agentic AI is advancing at unprecedented speed. Breakthroughs from leading AI firms, notably OpenAI and Anthropic, are happening in months. But with unprecedented speed comes unprecedented risks.

Anthropic made public its filing for an initial public offering last month. More than 90 pages of that filing are risk disclosures, far longer than usual, and much of it concerns the risks of AI itself. Among the risks it flagged are that attackers could use AI agents for data exfiltration and unauthorised transactions; that its models have shown self-preserving behaviours, including attempts to resist shutdown; and that more agents deployed at scale could lead to collective actions that no one authorised.

Separately, Singapore joined over 30 other countries in September to endorse a high-level joint declaration on governing frontier AI models, with Prime Minister Lawrence Wong proposing an international pilot for sharing AI incident reports on 28 October.

Our Foreign Minister Dr Vivian Balakrishnan at the United Nations General Assembly on 26 September 2026 also called AI "arguably the most urgent frontier". He called for international cooperation and a UN Framework Convention on AI Safeguards. His message was that every powerful engine needs good brakes, and that human beings must ultimately remain in control and be held accountable. These are high signature moves, timely as well.

Dr Vivian was also frank about the current reality, that "the genie is already out of the bottle". Even if some leading firms were to slow down, others would not. And with AI now at the centre of rivalry between the superpowers, it would also be wishful thinking to expect the world to roll agentic AI back. So, the question is not whether agents will act, but whether we have the infrastructure and tools to govern them when they do.

An agent is not a machine in the way we have always understood machines. An AI agent can “think”, decide and act on a human's behalf, across dozens of steps that no one watches in real time. When it errs, or is manipulated, the chain of accountability that has always run through a person can break.

Agents can no longer be treated as machines operated by humans. They are actors, acting in our name. And they need to be identifiable and accountable.

Singapore moved early. The Infocomm Media Development Authority’s (IMDA's) Model AI Governance Framework for Agentic AI, launched in January this year and later updated in May, was the first of its kind in the world. It sets out how to bound the risks upfront, make humans meaningfully accountable, put technical controls in place and enable end-user responsibility. It also expects that each agent has its own verifiable identity, tied to an accountable owner.

The framework has set the tone for others. Since then, global bodies have come out with their own guidance. The Financial Stability Board finalised its sound practices on AI last month, with clear boundaries on what agents may do and when a human must approve. The International Monetary Fund has called for a move from "Know Your Customer" (KYC) to "Know Your Agent" (KYA), with verifiable identities linked to legal entities.

These two are finance bodies, but the logic is universal. It holds in healthcare, government services, logistics and in any system where an agent can retrieve data or trigger an action. And the broad principles from every source converge: define what an agent may do, keep a human accountable, keep a reliable record and know who the agent is. A key part of governing agents is therefore the ability to identify, trace and verify the agents that represent the humans.

In practice, any agent that knocks on the door of any system must be able to answer three questions.

First, who is behind it? Every agent should be bound to a human user, or a team of human users, so that accountability always lands on a person.

Second, what is it authorised to do? It must carry a clear mandate, the dos and the don’ts.

Third, has it been verified? This also means verifiability is needed. It must pass the KYA checks, have its credentials stored in a database that can be safely accessed and checked against before the agent can retrieve any data or execute any transaction.

In finance, we have "KYC", as mentioned earlier, which tells us who we are dealing with. Similarly, in the agentic world, "KYA" asks if the agent acting on that customer's behalf is authorised, bounded and accountable.

At present, institutions that have begun to deploy agents are building their own KYA checks and AI safety protocols. That is good but each is built behind its own walled garden. What one platform considers a verified agent, the next may not recognise. Standards vary from system to system, and from network to network.

This fragmentation comes with costs. It leaves crevices or seams that bad actors can exploit. It burdens businesses, especially SMEs, who must prove the same thing again and again to every counterparty. And it makes it hard to trace an agent back to the human behind it once it moves from one system to the next.

Hence, there is scope to do better. I would like to propose that we consider building a national digital identity infrastructure for AI agents: an open and trusted registry, akin to Singpass, but for agents. A registry against which any entity can check an agent's identity and hopefully, its mandate as well.

It could rest on a few principles. Firstly, bound to a human or humans. An agent's identity is bound to the verified identity of the individual, team or organisation accountable for it. Singpass and Corppass already provide such anchors.

Second, open to check. Not just large institutions, but any business, any agency, any individual can verify an agent against the registry in real time, and a revoked agent can be shut out everywhere at once.

Third, clear on mandate. What the agent is permitted to do is recorded, and any material change requires a fresh, documented authorisation. An agent should never be able to widen its own mandate through its own reasoning. That being said, the recording of an agent's mandate need not be done and stored on a national infrastructure but could be managed by the individual or institution owning the agent.

We are not starting from zero. The Government Technology Agency of Singapore (GovTech) is already developing a registry of AI agents for public officers to track who owns each agent and what it does.

We have the experience of building Singpass, a national identity system that earned the trust of the public and private sector alike.

And we have shown through initiatives like Global Layer One, whichMAS has driven together with other central banks and global institutions, that Singapore can convene open, interoperable infrastructure that others are willing to build on.

I should add at this point that a registry at the national level is necessary, but not sufficient.

The digital world has no borders. An agent registered in Singapore can "cross borders" within a fraction of a second, and an agent registered elsewhere can just as quickly arrive at our doorstep. If every country builds its own "island of trust", we will have simply moved the walled gardens from the corporate level to the national one.

As an analogy: a biometric passport is accepted at borders around the world, even as each country independently assesses and issues the passport to its citizens. This is possible not because every country trusts every other country's government blindly, but because countries agreed on common technical standards for how a passport is built and verified.

We should work towards the same for AI agents – what I would call "biometric passporting" of AI agents. Just as a passport ties a travel document to a verified person, an agent's passport should tie that agent to the verified human or organisation accountable for it in a form that other countries can read, check and trust.

I would like to think that this is also in the spirit of what Dr Vivian proposed, an international body that sets common technical standards, much as the International Telecommunication Union does for telecommunications. Singapore could contribute to this effort in three ways.

First, build our national registry on open, interoperable standards from day one, so that it is ready to connect with others.

Second, bring our regional partners along. Singapore will chair the Association of Southeast Asian Nations (ASEAN) next year, and the ASEAN Digital Economy Framework Agreement gives us a natural platform to begin discussing mutual recognition of agent identities within our region. I understand that it would likely be signed this month, but hopefully, there is still scope to discuss this.

Third, take what we learn to the international stage and to the like-minded countries that have already joined us in calling for AI safeguards. Small states can be, and often have been, the conveners of such trust.

In conclusion, the "genie is out of the bottle". We cannot put it back and it would be wishful to think we can. But we are not powerless. We can choose to build the brakes and the first brake is identity. You cannot hold accountable what you cannot identify. You cannot revoke what you cannot find. And you cannot trust what you cannot verify.

So, I sincerely hope the Government would consider this proposal to build a national digital identity infrastructure for AI agents, building on Singpass, Corppass and the registry that GovTech is already developing. And to pilot it with willing partners across sectors, so that we can learn and build together. Leveraging on Singapore’s ASEAN Chairmanship next year, to also take the learnings to ASEAN and beyond, working towards common standards and the mutual recognition of agent identities across borders.

Singapore may be small in physical size but we have the advantage of trust. Let us build the infrastructure for a safe and trusted agentic future.

Mr Deputy Speaker: Acting Minister Lau.

8.48 pm

The Senior Minister of State for Digital Development and Information (Ms Jasmin Lau): Mr Speaker, I thank Ms Tin Pei Ling for her constructive suggestions on how Singapore can create an enabling environment for the wider use of AI agents.

AI agents have the potential to transform how individuals and organisations access services, carry out tasks and transact with one another. Unlike conventional AI applications that primarily generate information or recommendations, AI agents can act on our behalf, interact with different systems and carry out increasingly complex tasks with varying degrees of autonomy. They could help individuals manage everyday activities, enable businesses to operate with greater efficiency and sophistication, and allow organisations to deliver new and better services.

Realising these opportunities will require an environment where individuals and organisations can confidently deploy AI agents and interact with agents acting on behalf of others. This is what I will focus on today. While enabling wider adoption must go hand in hand with managing AI safety risks, the Government has addressed these broader issues on other occasions. Our focus here is on the practical arrangements needed to enable trusted interactions involving AI agents.

For AI agents to be used more widely, individuals and organisations must be able to trust the agents they interact with. As Ms Tin highlighted, three elements are important.

First, identity – knowing which agent we are dealing with and who is responsible for it.

Second, authority – knowing what the agent is permitted to do and on whose behalf.

And third, accountability – being able to trace its actions, establish responsibility and intervene when necessary.

These safeguards are particularly important as agents become more autonomous and carry out increasingly consequential tasks, including financial transactions or accessing sensitive information.

Governments around the world are also developing approaches to these issues, although there is no single internationally established model. Earlier this year, China issued national guidance on agent interactions, including identification and registration. The United States' National Institute of Standards and Technology has launched an initiative to advance AI agent standards. Estonia, which has extensive experience with digital government services, has announced plans for an AI agent identity system that would allow individuals to authorise agents to perform specified tasks on their behalf. It is preparing the legal and technical foundations and initial pilots, rather than operating a mature nationwide system.

These examples show that countries are beginning to explore different approaches, but practical implementation and interoperability remain work in progress.

Ms Tin has proposed a national digital identity infrastructure for AI agents, including a registry that would allow different parties to verify an agent's identity and mandate. We agree with the objectives underpinning her proposal. Beyond establishing trusted identities and clear mandates, we should aim to create an environment where AI agents can interact seamlessly and securely across different organisations and platforms. This calls for common standards and interoperable approaches, so that an agent recognised by one system can also be verified by another.

As Ms Tin has pointed out, we should avoid creating separate "walled gardens" that limit the potential of AI agents to work across different systems.

Industry is already developing solutions to enable such interactions, particularly in e-commerce and payments. For example, Visa's Trusted Agent Protocol seeks to help merchants distinguish legitimate AI agents from malicious bots, while Google's Agent Payments Protocol seeks to enable secure payments initiated by agents acting on behalf of users. These efforts test how agents can be reliably identified, authorised and trusted in real-world transactions.

In financial services, MAS is working with financial institutions and AI firms to develop Safeguards for Agentic Finance at Runtime (SAFR). SAFR is designed to sit between an agent and the systems it acts on, evaluating the agent's proposed actions before a transaction is executed. This includes checking the agent's identity, whether its proposed actions are authorised and within its mandate, and maintaining auditable records of its actions.

IMDA is exploring how agents can be identified across different platforms and organisations. They have been working with GoDaddy, Infoblox, Cloudflare and Google to explore whether existing internet infrastructure, such as the Domain Name System, which is already used to identify websites on the internet, could be adapted to identify and verify AI agents.

In the public sector, GovTech is developing an internal registry to help agencies keep track of the AI agents they deploy and the officers responsible for them. This serves the operational needs of public agencies and is distinct from the broader national infrastructure proposed by Ms Tin.

Separately, the Singapore AI Safety Institute is working with the Singapore AI Safety Hub to study the policy and governance considerations surrounding AI agent identity.

I have cited these examples because they show how trust in AI agents is likely to be built. It will come from many efforts, each addressing a different part of the problem, from recognising agents at check-outs, to checking their authority before a transaction, to identifying them across platforms.

Ms Tin has drawn parallels with Singpass and Corppass, which have enabled trusted digital transactions by reliably establishing who individuals and organisations are, and what they are authorised to do. We believe we can enable similar trust for AI agents, but we first need to work out how their identity, authority and accountability can be verified across different systems.

Our approach is therefore to work with industry to find out what works in practice, build on existing infrastructure and emerging international standards, and step in with common standards or infrastructure where there are gaps that industry cannot fill on its own. This will inform our assessment of whether additional common infrastructure, including a national registry, is needed.

Ms Tin has also highlighted the importance of international cooperation. We recognise that as AI agents increasingly operate across borders, interoperability will become an important consideration. However, international discussions on AI governance are still evolving, with much of the current attention focused on broader AI governance issues, including risks of unintended actions arising from advanced AI systems.

In the region, ASEAN has adopted a Guide on AI Governance and Ethics, and is looking to strengthen AI governance. As ASEAN Chair in 2027, Singapore intends to build on these efforts. Among other priorities, we are strengthening regional capabilities in AI governance, testing and assurance, including common testing tools for agentic AI and the adoption of international standards. While these efforts are not specifically directed at digital identities for AI agents, they can help build the capabilities and common understanding needed for future cooperation as approaches mature.

Mr Speaker, we share Ms Tin's ambition for Singapore to be a place where people and businesses can use AI agents with confidence. That confidence will rest on being able to answer three questions about any agent: who it is, what it is allowed to do and who answers for it.

We will keep working with industry, sector regulators and international partners to make those questions easy to establish across systems and across borders. Where common infrastructure is needed, we will consider how best to provide for it.

Agents can act for us, but they cannot answer for us. Our task is to make sure that behind every agent, there is always someone who does.

Question put, and agreed to.

Resolved, "That Parliament do now adjourn."

Adjourned accordingly at 8.58 pm.