What testers will probe

Testers may conduct adversarial “red-team” assessments, including attempts to prompt-inject or jailbreak systems—i.e., try to circumvent model safeguards and induce policy-violating or otherwise unsafe behaviour.

The assessment scope can cover:

• Safety and reliability failures, including inaccurate or fabricated outputs and harmful or otherwise undesirable content.
• Fairness failures, including bias in decision-making.
• Privacy and security problems, including data leakage and vulnerabilities.
• Prompt-injection and jailbreak resilience, particularly relevant to generative and agentic AI applications.

Accreditation requirements

Applicants will be assessed on both their testing capability and their organisational capacity:

• Technical competence: evidence that the firm can perform its claimed specific testing services, assessed against guidance such as IMDA’s Starter Kit for Testing LLM-Based Applications for Safety and Reliability.
• Governance and business standing: relevant track record, good standing and a commitment to business activity in Singapore.
• Financial sustainability: sufficient financial standing to operate sustainably.
• Operational readiness: adequate operational capacity and readiness to provide the proposed services.
• Accurate business scope: the firm’s declared service scope must match its demonstrated capabilities.

The Starter Kit is voluntary guidance for baseline LLM safety and reliability testing. Its five stated risk areas are:

1. Hallucination and inaccuracy.
2. Bias in decision-making.
3. Undesirable content.
4. Data leakage.
5. Vulnerabilities.

Fees, duration and early interest

• AI Verify Foundation says there will be no application or accreditation fees, to keep participation barriers low and support the AI-assurance market’s growth.
• Accreditation will be valid for 12 months, after which firms must renew and demonstrate continued competence as standards and the threat landscape evolve.
• Firms reported as having expressed early interest are Advai, AIDX, Ernst & Young, Knovel Engineering, PwC, Resaro and Vulcan.

Expected industry effect

The intended benefit is a more consistent and credible external-testing market: common expectations for tester competence, evidence-based and documented assessments, and testing tailored to the organisation’s AI use case and risk scope.

Industry representatives’ specific comments about improved customisation, governance or confidence were not included in the available evidence. However, the programme’s stated design supports those outcomes by helping buyers identify capable testers, making the test scope explicit, and giving deployers additional assurance from independent assessment.

What testers will probe

Testers may conduct adversarial “red-team” assessments, including attempts to prompt-inject or jailbreak systems—i.e., try to circumvent model safeguards and induce policy-violating or otherwise unsafe behaviour.

The assessment scope can cover:

• Safety and reliability failures, including inaccurate or fabricated outputs and harmful or otherwise undesirable content.
• Fairness failures, including bias in decision-making.
• Privacy and security problems, including data leakage and vulnerabilities.
• Prompt-injection and jailbreak resilience, particularly relevant to generative and agentic AI applications.

Accreditation requirements

Applicants will be assessed on both their testing capability and their organisational capacity:

• Technical competence: evidence that the firm can perform its claimed specific testing services, assessed against guidance such as IMDA’s Starter Kit for Testing LLM-Based Applications for Safety and Reliability.
• Governance and business standing: relevant track record, good standing and a commitment to business activity in Singapore.
• Financial sustainability: sufficient financial standing to operate sustainably.
• Operational readiness: adequate operational capacity and readiness to provide the proposed services.
• Accurate business scope: the firm’s declared service scope must match its demonstrated capabilities.

The Starter Kit is voluntary guidance for baseline LLM safety and reliability testing. Its five stated risk areas are:

1. Hallucination and inaccuracy.
2. Bias in decision-making.
3. Undesirable content.
4. Data leakage.
5. Vulnerabilities.

Fees, duration and early interest

• AI Verify Foundation says there will be no application or accreditation fees, to keep participation barriers low and support the AI-assurance market’s growth.
• Accreditation will be valid for 12 months, after which firms must renew and demonstrate continued competence as standards and the threat landscape evolve.
• Firms reported as having expressed early interest are Advai, AIDX, Ernst & Young, Knovel Engineering, PwC, Resaro and Vulcan.

Expected industry effect

The intended benefit is a more consistent and credible external-testing market: common expectations for tester competence, evidence-based and documented assessments, and testing tailored to the organisation’s AI use case and risk scope.

Industry representatives’ specific comments about improved customisation, governance or confidence were not included in the available evidence. However, the programme’s stated design supports those outcomes by helping buyers identify capable testers, making the test scope explicit, and giving deployers additional assurance from independent assessment.