Singapore’s Infocomm Media Development Authority has rolled out three initiatives aimed at making AI systems safer and more resilient as agentic tools move into wider use. The measures, announced on October 14, 2026, at the AI Assurance Exchange: Agentic Edition, focus on practical testing guidance, independent testers, and sector-level defenses against AI-enabled cyber threats.
IMDA chief executive Ng Cher Pong said the work responds to a basic shift: “When we give AI more agency, we need greater assurance over how that agency is exercised.” The initiatives build on Singapore’s existing Model AI Governance Framework for Agentic AI and earlier guidance for testing large-language-model applications.
Starter Kit for Testing Agentic Systems
The first measure is the Starter Kit for Testing Agentic Systems for Safety and Reliability. It gives organizations a four-step process for assessing AI agents before and after deployment: identify risks such as erroneous or unauthorized actions, design realistic tests, run the tests and analyze logs, and continuously monitor and re-test systems.
Unlike simpler LLM applications, agentic systems can use tools, access sensitive data, and take actions on a user’s behalf. The kit therefore calls for examining an agent’s full trajectory—its reasoning, the tools it invokes, the information it reaches, the decisions it makes along the way, whether it stayed within granted permissions, and whether safeguards performed as expected—rather than judging only the final output.
IMDA developed the kit after consultations with government agencies and more than 20 industry deployers and testers, including Amazon, Google, Microsoft, and Singtel. It also draws on testing experience from the Global AI Assurance Sandbox and the Singapore AI Safety Institute. The authority has said the kit will be updated as methods and best practices evolve.
AI Tester Accreditation Programme
The second initiative is the AI Tester Accreditation Programme, launched by the AI Verify Foundation, a nonprofit wholly owned by IMDA. The program evaluates third-party testing firms on technical competence, operational readiness, financial sustainability, and track record. Accreditation is granted for specific testing methods (such as benchmarking and red-teaming) and specific risk areas (including hallucination, undesirable content, data leakage, vulnerability to adversarial prompts, and bias and fairness). It is valid for one year.
Five firms have already been accredited: AIDX Tech, Knovel Engineering, PwC Singapore, Resaro, and Vulcan. The credential signals demonstrated capability in defined areas; it does not certify that any particular system a firm tests is safe, nor does it cover every type of AI system. IMDA has noted that organizations increasingly view accredited testers as more credible sources of independent assessment.
Frontier AI Resilience and Security Taskforce
The third measure is the creation of the Frontier AI Resilience and Security Taskforce (FAIRS). It brings together more than 20 leaders and experts, primarily based in Singapore, from the telecommunications, media, cybersecurity, and AI sectors. The taskforce will focus on three priorities: improving the speed of responses to AI-accelerated threats, strengthening critical systems as AI becomes more effective at finding vulnerabilities, and pooling intelligence across the sector so emerging attack techniques can be identified earlier.
FAIRS is expected to publish a report with practical guidelines and best practices in the first half of 2027. The effort sits alongside a broader national cybersecurity strategy that Singapore launched around the same time, which includes plans to use AI more actively in national cyber defense.
Taken together, the three initiatives aim to give organizations clearer methods for testing agents, a more reliable market of independent testers, and a coordinated industry response to the cyber risks that frontier models introduce. IMDA has emphasized that no single safeguard is sufficient; the package is designed to address both the behavior of AI systems themselves and the ways those systems can be used by attackers.