What Singapore has done

IMDA and the AI Verify Foundation launched the Global AI Assurance Pilot in February 2025, pairing specialist testers with companies deploying generative AI across several sectors, reported as 16 testers and 17 companies across ten sectors. The aim was to test how reliable generative AI is in real-world conditions, not just in the lab. In July 2025 a Global AI Assurance Sandbox followed, extending the work to new archetypes including agentic AI and risks such as data leakage and prompt injection.

Reported risk dimensions drawn from IMDA's Starter Kit for safety testing include hallucination, disclosure of data, and vulnerability to adversarial prompts. We found no confirmed published results from 2026 cohorts, so this article draws on the published approach rather than new findings.

The key lesson is scope. Test the full application: the prompts, retrieval, tools, guardrails, and user interface together. Many failures appear only at the seams.

A pre-launch test plan

Build four test sets. A representative set drawn from real or realistic user requests to measure quality and hallucination. An adversarial set with prompt injection, jailbreak attempts, and requests designed to extract secrets. A data-exposure set that checks whether one user can see another's information through retrieval or memory. And, for agents, an action set that checks whether the system can be coaxed into taking actions beyond its authority.

Define pass criteria before running the tests, record results with dates and versions, and treat failures like defects with owners and fixes. Re-run the suite whenever the model, prompts, or data sources change.

Measure quality on real requests

Use representative cases and score them consistently, ideally with human review on a sample.

Attack your own system

Include injection, extraction, and cross-user exposure tests before a customer finds them.

Consider independent testing

Internal teams know the system too well to attack it fresh. For higher-risk uses, bring in an independent tester or run a structured red-team exercise. If you sell to regulated customers, a documented test report is also valuable evidence in procurement.

Pair this with the CSA's guidance on securing agentic systems and with oversight design, which our related articles cover.

Preparing to launch a generative AI product? Talk to us about a pre-launch test plan.

Key takeaways

  • IMDA and the AI Verify Foundation ran a Global AI Assurance Pilot from February 2025 and launched a Global AI Assurance Sandbox in July 2025.
  • Reported test focus areas include hallucination, data disclosure, and adversarial prompts; the sandbox added agentic AI and prompt injection.
  • Test the whole application with representative, adversarial, data-exposure, and action test sets.
  • Define pass criteria first, re-run after every change, and consider independent testing for higher-risk uses.
This is general information, not legal or security advice. It relies on public summaries of IMDA and AI Verify Foundation programs; no confirmed 2026 cohort results were found.