How Singapore Is Scaling AI Across Government Without Giving Agents Free Rein

Singapore is taking a systems-level approach to public-sector AI: expand access while putting enforceable controls around the agents, tools and data they can use. GovTech’s AI Assistant Desk and planned AI-agent registry are designed to make government AI activity visible and accountable as adoption scales to about 150,000 public officers.

The initiative is still being developed and piloted with selected officers, with a broader rollout planned for later in 2026. Its significance is less about introducing another chatbot than about creating a common operating layer for AI governance across government.

The AI Assistant Desk is being built around accountability

The registry is intended to record who owns an AI agent, what it is allowed to do and how it is used. That gives agencies a way to connect an automated system to a responsible owner and review its activity over time.

This matters because an agent can do more than generate text. Depending on its permissions and integrations, it may make decisions, call tools or carry out tasks at machine speed. A registry therefore functions as an inventory and accountability mechanism for digital workers—not just a catalogue of software.

GovTech has described the wider Assistant Desk as a secure digital-assistant environment for public officers. The expected use cases include drafting reports, managing schedules, coding, research and other routine work.

Technical guardrails define what agents can do

The proposed model is to enforce safety through the surrounding system rather than relying only on the safeguards built into an individual AI model. Reported controls include:

• configurable rules for different agencies or use cases;
• approved or sanctioned AI tools;
• restrictions on actions such as deleting files or sending external email;
• limits on recipients to reduce the risk of uncontrolled distribution or spam; and
• automated checks on prompts and outputs for offensive language and other hygiene issues.

These controls reflect a least-privilege principle: an agent should receive only the access and action space required for its assigned task. That approach is especially important when agencies use multiple models or connect agents to sensitive government systems.

The safety layer is intended to remain useful when models change

GovTech’s AI Guardian provides a separate governance layer for AI applications. Its two named components serve different stages of the lifecycle: Litmus assesses safety and security risks before deployment through adversarial testing, while Sentinel provides real-time content moderation and output filtering once an application is live.

Separating these functions from the underlying model can reduce dependence on any single provider. GovTech has also described the Assistant Desk’s security architecture as designed to preserve its protections when third-party AI tools are added or replaced.

The distinction is important: a model may change, but the government’s requirements for data protection, permitted actions, monitoring and output safety should remain stable. GovTech’s responsible-AI framework identifies safety, robustness, security and privacy, explainability and transparency as core requirements.

Adoption is paired with capability-building

Singapore’s public-sector AI programme is not limited to autonomous agents. GovTech’s Pair assistant already supports writing, research, summarisation and analysis on government-issued devices, including work involving sensitive government data that commercial chatbots cannot handle.

GovTech is also developing AI-assisted software engineering and low-code approaches so technical teams and public officers can prototype or automate workflows more quickly. Earlier GovTech reporting described plans for AI code assistants, AI-supported engineering and tools intended to help agencies move prototypes into safer environments.

Training is another part of the infrastructure. GovTech provides data and AI learning resources for public officers, while Singapore’s wider workforce plans include AI-skills programmes for technology professionals working on automated coding and agentic systems.

The practical objective is to build internal capability: officers need to know when AI is appropriate, technical teams need to evaluate and secure what they build, and agencies need shared platforms rather than a collection of disconnected experiments.

Public-sector projects show why evaluation matters

GovTech’s work on LangBuddy illustrates the need to evaluate an AI application against its real purpose. LangBuddy is an AI voice chatbot designed to help students practise their Mother Tongue language through conversation.

For that kind of product, a useful evaluation cannot stop at whether the chatbot produces fluent replies. GovTech’s evaluation work considers safety, hallucination risk and whether responses follow the intended pedagogical rules.

That principle generalises across public services: agencies need to test whether an AI system performs the task it was designed for, follows its constraints and behaves acceptably in realistic interactions—not merely whether it sounds convincing.

Cybersecurity is both a prerequisite and an AI use case

The same governance philosophy appears in GovTech’s cybersecurity work. Its stated capabilities include secure system design, red-team exercises, penetration testing, crowdsourced vulnerability disclosure and blue-team functions for detecting and responding to threats.

GovTech is also testing AI-augmented penetration testing and automated security-operations triage. These trials place AI on both sides of the security process: helping discover weaknesses and helping defenders prioritise incidents.

Some reporting refers to more than 2,000 government digital systems in the context of scaling security testing. The available material supports the broader direction—red-teaming, penetration testing and AI-assisted security operations—but does not establish that continuous AI testing is already running across every one of those systems.

What the programme means for Singapore’s wider strategy

Taken together, the initiatives support three connected goals:

1. Improve public-service delivery. Assistants and agents can handle drafting, analysis, coding and routine workflows, giving officers more capacity for complex and citizen-facing work.
2. Strengthen in-house technology capability. Shared platforms, evaluation tools, security layers and training help Singapore retain operational knowledge instead of depending entirely on one model provider or commercial product.
3. Protect public trust and national systems. Registries, permissions, auditability, safety testing and cyber defence create oversight around AI as it gains access to government work and data.

The central lesson is that scaling AI safely is an architecture problem as much as a model problem. Singapore is attempting to make AI broadly available while keeping ownership, permissions, testing and activity monitoring visible to the institutions responsible for public services.