Singapore's digital ministry told Parliament it favours risk-based safeguards for frontier and agentic AI and does not require pre-deployment model access

Key takeaways

• The reply is non-binding. It answers four Members' questions in one written answer and announces no new requirement.
• On pre-deployment access, MDDI says "We currently do not make this a requirement, because access to a model by itself does not necessarily give a complete picture of the risks."
• For AI agents, the Government says it limits what systems and data an agent can reach and what actions it can take, and points the wider economy to IMDA's agentic AI framework.
• For essential services and other higher-risk uses, the reply says stronger safeguards are necessary and that the Government will continue to study the need for new requirements.

Is the MDDI reply binding, and what did the ministry actually say?

No. It is a written answer to Parliamentary questions, published on the MDDI newsroom. It responds to Question 33 from Mr Fadli Fawzi, which asked whether AI companies give the National AI Council or another agency their latest frontier model before deployment for independent evaluation, and whether the Government will seek such access. The minister's answer also covers questions from Ms Valerie Lee, Dr Charlene Chen and Mr Alex Yam, and related questions filed by three further Members for later sittings. The reply sets out four parts: set limits based on risk, strengthen defences against unintended actions and misuse, detect risks and respond early, and work with international and industry partners. It does not create a duty, a deadline or a penalty. Our article on the MDDI supply chain deployer duty covers a separate instrument.

What does the reply say about AI agents and risk-based limits?

Safeguards should scale with harm. The reply says limits should reflect the sensitivity of the data and systems involved, the degree of autonomy given to the AI system, and the severity and remediability of any potential harm, and that "The greater the potential harm, the stronger the safeguards should be." It describes how the Public Service approaches AI agents: what systems and data an agent can access, what actions it is allowed to take, and the consequences of unintended or unauthorised actions, complemented by policies and governance processes for oversight and accountability. For the broader economy it points to IMDA's Model AI Governance Framework for Agentic AI and Starter Kit for Testing LLM-Based Applications, which it says provide guidance on managing risks and maintaining meaningful human accountability, and to the Global AI Assurance Sandbox for real-world testing of generative and agentic AI risks. The reply also says that telling agents what to do is not enough: it lists limiting what they can access and do, test suites, human oversight, monitoring and containing consequences.

For essential services and other higher-risk applications, the reply says stronger safeguards are necessary. It gives as examples more rigorous testing, independently verifiable evidence that safeguards are effective, tighter deployment controls or tighter regulatory oversight, and says the Government will continue to study the need for new requirements, especially in high-risk uses. These are examples the reply raises, not announced rules.

Does Singapore require pre-deployment access to frontier models, and how will it detect AI incidents?

Not at present. The reply says the Government engages frontier developers to obtain information and access to their models where appropriate, but does not make this a requirement. It gives two reasons: access alone does not necessarily give a complete picture of the risks, and a rigid requirement could be counter-productive if it leads companies to limit cooperation or information-sharing. It prefers collaborative relationships with frontier labs and says it draws on developers' own assessments, independent research and evaluations, and the work of other AI safety institutes and international partners.

On incidents, the reply notes that established channels through CSA, GovTech and sector leads already receive cybersecurity incident reports, including those involving AI. It says the Government is looking into how these channels can better identify AI incidents and whether further coordination or reporting arrangements are needed. Responding to Mr Fadli Fawzi, it states that no agency has received a report of a cyber attack against its systems involving an unsupervised AI agent so far. The Singapore AI Safety Institute is described as building technical capabilities to evaluate advanced AI systems.

What should counsel and compliance leads do now?

Treat the reply as a signal about direction, not a compliance standard. Nothing in it applies a new duty to your organisation, and the reply itself says the study of new requirements is continuing.

• Compare your AI agent permissions, testing and human-oversight points with the controls the reply describes and with IMDA's agentic AI framework.
• Map which of your systems could count as essential services or other higher-risk uses, since the reply points to stronger safeguards there.
• Check how an AI-related incident would be reported internally and to the existing CSA, GovTech or sector channels the reply mentions.
• Keep a dated record of your position so you can show how it changes if requirements are later proposed.

What are the key dates and parties?

What we did not verify

What we opened: the full text of the MDDI newsroom page containing the written answer, on 2026-10-25 (UTC).

What we did not open: the IMDA Model AI Governance Framework for Agentic AI, the Starter Kit for Testing LLM-Based Applications, the Global AI Assurance Sandbox materials, the international call for stronger safeguards the reply says Singapore endorsed, the Hansard record of the sitting, and any press coverage of the exchange. We rely on the reply's own description of each.

What we refuse to claim: that the reply is binding, that new agentic AI rules are coming, that any named sector will face tighter oversight, what other governments require, or anything about the reply beyond its text.

Related briefings

• AI Regulation News hub
• Asia-Pacific AI regulation tracker
• Attorneys and AI regulation
• MDDI and the AI supply chain deployer duty
• MAS runtime safeguards for agentic AI in finance
• CSA code of practice for critical information infrastructure
• IMDA generative AI chatbot transparency information card