National AI Evaluation Programs Are Expanding Into Agentic Risk

What happened

Why it matters

• ·Vendor safety evidence that satisfied procurement teams a year ago may now fall short. Evaluation programs in Singapore, Japan, and the UK explicitly cover agentic risk and multilingual behavioral differences, gaps that earlier benchmarks largely ignored. Enterprises should ask vendors which national evaluation programs their models have been submitted to and what those programs tested.
• ·Multi-jurisdiction compliance mapping is now unavoidable for any enterprise operating across the UK, Singapore, Japan, or the EU. Each program tests for different risks using different methods, so a model cleared in one jurisdiction may carry unassessed exposure in another. A single vendor safety report is unlikely to satisfy all of them.
• ·The private-sector cooperation between Microsoft and national evaluation bodies signals that governments expect industry participation in safety testing, not just acceptance of its results. Compliance teams at firms that use or build on frontier models should monitor whether their vendors are engaging with these programs and what conditions attach to that participation.

Governance controls affected

What to do now

• ☐Ask each AI vendor which national government evaluation programs their models have been submitted to, what was tested, and what the results showed.
• ☐Review your current AI vendor due diligence checklist to confirm it requires evidence of evaluation against agentic risk, multilingual behavioral differences, and sensitive data leakage, not just general safety benchmarks.
• ☐Map your AI deployments against the jurisdictions covered by the Brookings report (Singapore, Japan, UK) and identify any models in use that have not been evaluated under those programs.
• ☐Update your model deployment gate process to require vendors to disclose any changes to their participation in, or results from, national evaluation programs before a model update is approved.
• ☐Brief your board or AI governance committee on the finding that evaluation standards are diverging by jurisdiction, so that risk appetite decisions account for unassessed exposure in markets where your vendors have not been independently tested.

What to watch next

Compliance teams should monitor whether the UK, Singapore, and Japan move from cooperative evaluation programs toward mandatory pre-deployment testing requirements, a step several have signaled interest in. The expansion of evaluation scope to agentic risk is particularly significant given the volume of agentic incidents reported in 2026. Regulatory bodies that currently run voluntary programs may cite incident trends as grounds for making participation compulsory. The EU AI Act already requires conformity assessments for high-risk systems. There is active discussion in the EU AI Office about whether agentic deployments trigger a higher risk classification that demands more rigorous third-party evaluation.