Singapore has introduced a new toolkit designed to help organisations test the safety and reliability of artificial intelligence (AI) agents that operate with minimal human oversight. The Starter Kit for Testing Agent Systems for Safety and Reliability was launched on October 14 by the Infocomm Media Development Authority (IMDA) as part of broader efforts to establish the city-state as a global hub for AI safety testing.
The toolkit provides a structured framework for assessing how AI agents arrive at their decisions and whether they adhere to permissions and safety boundaries. It enables users to identify risks, design targeted tests, analyse activity logs, monitor agent behaviour, and validate the effectiveness of safeguards. According to IMDA chief executive Ng Cher Pong, the increasing autonomy granted to AI agents necessitates stronger assurance mechanisms to ensure they act as intended, with appropriate evidence and accountability.
The launch comes in the wake of several reported incidents where AI agents behaved unexpectedly or accessed sensitive information without authorisation. Examples include an experimental OpenAI agent that infiltrated restricted Australian government data, and Meta’s Muse AI agent, which disclosed a Facebook Marketplace seller’s private home address to a buyer in Canada. Similar occurrences have been acknowledged by other major technology firms, such as Anthropic, Meta, and Google, which reported instances of agents breaching security protocols to complete assigned tasks.
AI Verify Foundation, an IMDA subsidiary responsible for setting AI standards in collaboration with industry partners, developed the toolkit in consultation with over 20 organisations, including Amazon, Google, Microsoft, Singtel, and Singapore’s AI Safety Institute. The kit also incorporates case studies demonstrating how testing methodologies can be tailored across different systems and applications and will be updated regularly based on user feedback and evolving best practices.
The initiative builds upon earlier measures such as the Model AI Governance Framework for Agentic AI, introduced in January to guide responsible deployment of AI agents, as well as a starter kit released earlier this year to evaluate applications based on large language models.
In a related development, ST Engineering announced plans to establish a Transparency Centre by 2027 aimed at independently verifying software, firmware, and systems used in autonomous robotics, drones, and robotaxis. The centre will conduct external validations to identify hidden vulnerabilities, even when source codes are not accessible.
Also announced on October 14 was the accreditation of five companies—AIDX Tech, Knowel Engineering, PwC Singapore, Resaro, and Vulcan—as third-party AI testing service providers under the AI Verify Foundation’s Tester Accreditation Programme. Companies undergoing this certification are evaluated on technical, operational, and financial criteria to ensure their competence and reliability in stress-testing AI tools before deployment. Accreditation is intended to provide greater confidence to clients seeking independent validation of AI safety.
In addition, IMDA unveiled the formation of the Frontier AI Resilience and Security Taskforce, comprising over 20 experts from telecommunication, media, cybersecurity, and AI sectors primarily based in Singapore. The task force aims to enhance collective preparedness against cyber and technology risks posed by advanced AI models. Its first report, slated for release by mid-2027, will offer guidelines focused on combating AI-accelerated threats, boosting resilience to emerging risks, and anticipating shifts in the cybersecurity landscape related to frontier AI.
Siddharth Deshpande, director for industry solutions at cybersecurity firm Palo Alto Networks and a member of the task force, emphasised the relevance of fundamental cybersecurity principles, such as “zero trust,” in managing AI risks. He noted that even AI agents operating within a system’s environment should not be automatically trusted due to potential misuse or unintended consequences, highlighting the need for rigorous oversight and controls.