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

Singapore launched a new national Cybersecurity Strategy at the Singapore International Cyber Week (SICW) on 13 October, putting artificial intelligence at the centre of how the country plans to defend itself. The same week, the Cyber Security Agency of Singapore (CSA) and Google ran an exercise on a newer problem: how to keep autonomous AI agents under control.

Three shifts

Announcing the strategy, K Shanmugam, Coordinating Minister for National Security and Minister for Home Affairs, described three shifts towards a more proactive defence: "Singapore will seek to act earlier, act at national scale, and act at the speed of the threat."

• Earlier: continuous monitoring, threat hunting and fixing weaknesses before adversaries exploit them.
• At national scale: coordinated responses across government and industry.
• At the speed of threats: using AI in defence.

The strategy has four areas: strengthening critical infrastructure and the wider cyberspace, building capabilities in AI and quantum technologies, deepening international partnerships, and investing in the cybersecurity ecosystem and workforce. No funding figures were announced.

AI on the defending side

The government says it will progressively deploy AI across national cyber defence, including in a National Cybersecurity Command Centre (NC3), CSA's operations hub, and the Government Cyber Security Operations Centre. It will also invest in what it calls AI harness engineering: the ability to evaluate, choose between and securely connect the best available AI models to defenders' tools, data and workflows, rather than building a model for every task.

Some of this is already in use. CSA has built 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. A Protective Domain Name Service, which blocks connections to known malicious domains, is meant to cut exposure to phishing, scams and malware.

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

Keeping AI agents in check

That second half of the problem was the subject of a tabletop exercise that CSA and Google held on 14 October with senior cyber policymakers, critical infrastructure operators and AI safety experts. According to CSA's outcome statement, participants identified risks that grow as agents take over core functions at machine speed: unintended agent behaviour that is hard to detect, compromised agents, unexpected coordination between multiple agents, and over-reliance on a single model for highly privileged tasks such as security monitoring. Perimeter defences and manual human review, they noted, may leave gaps or become bottlenecks.

The safeguards they pointed to sit outside the model, rather than assuming models 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, combined with monitoring of agents' reasoning and tool use;
• independent verification and use of more than one model;
• verifiable agent identities based on standard protocols;
• incident response plans written with agents in mind.

Participants called for trusted channels to share threat intelligence and reports of unexpected agent behaviour, more stress-testing and pilot testing of agents before wide deployment, and continued human oversight of consequential decisions. CSA has also said it will develop guidance on securing autonomous robotic systems across their lifecycle.

Why it matters beyond Singapore

Singapore's frameworks tend to travel across the region, and the questions raised at SICW are the ones every organisation deploying AI agents will face. October has already shown the other side: PHP2Shell was found with AI tools and turned into a working exploit faster because of them. Defenders are being asked to use the same tools, and to keep their own agents on a short leash while they do.