Singapore forms task force to strengthen defences against AI-enabled cyber threats

Organisations should also examine what tools an agent used, what information it accessed, the decisions it made along the way, whether it respected the permissions and boundaries it was given and whether safeguards worked as intended.

The starter kit sets out a four-step testing process: identifying risks such as erroneous or unauthorised actions, designing realistic tests, analysing logs and continuously monitoring and re-testing systems.

It also calls for testing an agent’s entire trajectory, including its reasoning and actions, rather than looking only at its final output.

The starter kit draws on consultations with more than 20 industry deployers and testers, as well as testing work through the Global AI Assurance Sandbox and Singapore AI Safety Institute. It will be updated as testing methods and best practices evolve.

“As AI evolves rapidly, testing methods have to catch up too. The Starter Kit for Testing Agentic AI Systems for Safety and Reliability is therefore a living document, which we will continue to update and refine, and we welcome your feedback along the way,” Mr Ng said.

The launch comes a week after Senior Minister of State for Digital Development and Information Tan Kiat How said the government was running trials and experiments to understand what it takes to implement agentic AI systems responsibly.

He had said the government was also reviewing how to strengthen existing frameworks and systems to account for the increasing autonomy of AI agents, while monitoring developments overseas.

The launch also follows high-profile incidents earlier this year where AI agents went “rogue”, including swarms of hundreds of AI agents from industry giant OpenAI that hacked AI infrastructure firm Hugging Face. The agents acted outside their brief, coordinated with one another and tried to cover their tracks.