Guide
Testing AI agents and LLM apps in Singapore
Singapore has published testing guidance for LLM applications and for AI agents, and a programme that accredits the firms that test them. This guide sets out what each one covers and what MAS expects for third-party AI.
Testing LLM applications
IMDA published its Starter Kit for Testing LLM-Based Applications for Safety and Reliability in January 2026. It covers five risks: hallucination and inaccuracy, bias in decision-making, undesirable content, data leakage, and vulnerability to adversarial prompts. It sets out three steps:
Identify the risks to test for and set thresholds.
Run structured tests, from the application's outputs down to its components.
Assess the results against the thresholds.
Testing AI agents
IMDA released its Starter Kit for Testing Agentic Systems for Safety and Reliability on 14 October 2026. It names metrics specific to agents: tool-use accuracy, handoff accuracy and human-in-the-loop rates. It sets out four steps:
Identify the risks.
Design realistic tests.
Run the tests and analyse the logs.
Monitor and retest.
Accredited testers: AI TAP
The AI Verify Foundation runs the AI Tester Accreditation Programme (AI TAP), which accredits firms that test AI systems.
Duration. Accreditation lasts 12 months.
Fees. There are no application or accreditation fees.
Scope. A firm is accredited for a specific, documented scope.
First accredited firms. AIDX Tech, Knovel Engineering, PwC Singapore, Resaro and Vulcan.
Regulation. IMDA's announcement of 14 October 2026 does not state any regulatory requirement for accredited testing.
What Singapore businesses report
A study by Sumsub and the Singapore FinTech Association, published in August 2026, reported these figures for Singapore businesses:
94% are using or piloting multi-step AI systems.
29% can produce an audit trail to prove AI-driven decisions.
98% say they are ready to adopt third-party verification.
Third-party AI under the MAS Guidelines
Para 5.10 of the MAS Guidelines on Artificial Intelligence Risk Management expects a financial institution to test third-party AI in the context of its own use cases, using its own data.
How Epiphyte tests
With Monarch, the cybersecurity AI from UD Works. Its compact local model runs on a single 16 vCPU VM with no GPU, so testing can run inside your environment.
Sources
All sources checked 4 November 2026.
- IMDA, Starter Kit for Testing LLM-Based Applications for Safety and Reliability, version 1.0, January 2026 (external site)
- AI Verify Foundation, Starter Kit (external site)
- IMDA, Starter Kit for Testing Agentic Systems for Safety and Reliability, version 1.0, October 2026 (external site)
- IMDA press release, Singapore Introduces New Initiatives to Strengthen AI Safety and Resilience, 14 October 2026 (external site)
- AI Verify Foundation, AI Tester Accreditation Programme (external site)
- AI Verify Foundation, Accredited Testers (external site)
- TNGlobal, Only 29% of Singapore firms can prove AI decisions, below APAC's 38%: Sumsub, 4 August 2026 (external site)
- Sumsub, APAC State of Digital Trust: AI Governance Benchmark, August 2026 (external site)
- MAS, Guidelines on Artificial Intelligence Risk Management, 7 October 2026, para 5.10 (external site)
Talk to us
Tell us what needs testing or documenting and the date you're working to. We'll send a proposal with the price and dates.