Singapore Cybersecurity Strategy 2026: AI in Defence and Safeguards for AI Agents

The government unveiled Singapore's Cybersecurity Strategy 2026 at the opening of the 11th Singapore International Cyber Week (SICW) on 13 October, and artificial intelligence sits at the centre of how the country plans to defend itself. The next day the Cyber Security Agency of Singapore (CSA) and Google held a tabletop exercise on keeping autonomous AI agents under control.

Act earlier, act at national scale

K Shanmugam, Senior Minister, Coordinating Minister for National Security and Minister for Home Affairs, set out the change in his SICW opening address: "Singapore will seek to act earlier, act at national scale, and act at the speed of the threat."

Continuous monitoring and threat hunting are meant to find and fix weaknesses before attackers reach them. For national scale there's the National Cybersecurity Command Centre (NC3), CSA's operations hub, pulling threat information from government, critical infrastructure owners, security firms and international partners into one picture. Upstream, the government will help develop a national Protective Domain Name Service that blocks connections to known malicious domains before they reach the user.

Acting at the speed of the threat

Shanmugam said the government will progressively deploy AI across national cyber defence, including inside NC3 and the Government Cyber Security Operations Centre. It won't build its own model for every use case; he committed to what he called AI harness engineering instead: the ability to choose between AI models and securely connect the best available ones to defenders' tools, data and workflows.

"We cannot resist AI adoption. But we need to try and adopt it in a safe and secure manner," he said.

Some of this is already running. CSA has AI tools that automate the discovery and mapping of exposed assets and vulnerabilities, and GovTech uses AI agents to speed up vulnerability discovery and security testing. The first of the PHP2Shell flaws was found with an OpenAI model.

Keeping AI agents in check

CSA's outcome statement says the exercise on 14 October brought together senior cyber policymakers, critical infrastructure leaders and AI safety experts. Perimeter defences and manual human review alone can't keep up once agents run core functions at machine scale, which makes compromised agents, unintended behaviour and unexpected coordination between agents harder to spot. Where an agent does highly privileged work such as security monitoring, participants warned that relying on a single model is a risk in itself.

The safeguards they discussed sit "outside the model" rather than assuming a model will always behave as intended:

• continuous runtime monitoring, hardened execution environments and strict access controls;
• real-time, deterministic guardrails that block unauthorised high-impact actions, plus monitoring that audits agents' reasoning and tool use;
• independent verification;
• standard protocols for verifiable agent identities;
• incident response plans for fast-escalating agentic threats.

The ASEAN response

On 14 October, participants from all ASEAN member states met at the 11th ASEAN Ministerial Conference on Cybersecurity in Singapore. They welcomed the ASEAN Cybersecurity Cooperation Strategy 2026–2030 and encouraged regional cooperation on securing AI systems and using AI in cyber defence. Asked whether ASEAN was discussing coordinated responses to rogue AI agents, Shanmugam told reporters at a media doorstop on 13 October: "ASEAN has its own framework for dealing with cyber issues, but I would say it's a work in progress."