NIST CAISI Expands Frontier AI National Security Testing to Google DeepMind, Microsoft and xAI

NIST’s Center for AI Standards and Innovation (CAISI) announced on May 5, 2026 that it has signed expanded collaborative agreements with Google DeepMind, Microsoft and xAI for frontier AI model testing in the context of national security. The new agreements build on CAISI’s earlier contracts with Anthropic and OpenAI from August 2024, giving the US government formal evaluation arrangements with all five leading frontier labs in the United States.

What do the agreements specifically cover?

The agreements cover pre-deployment evaluations (before a model’s public release) and post-deployment research (analysis of models already on the market). CAISI has to date conducted more than 40 evaluations — including assessments of unreleased state-of-the-art models that labs submit for testing before launch.

The technical framework of the agreements allows labs to deliver models with “reduced or removed safeguards” (e.g., content filters, refusal layers), enabling CAISI to measure the true capability limits of models without interference from safety guardrails. Testing is routinely conducted in classified environments with interagency experts acting through the TRAINS Taskforce — a coordination body established in November 2024 to align AI research with national security.

How does the director’s statement shape the strategic context?

Chris Fall, CAISI’s director, summarized the purpose of the agreements: “Independent, rigorous measurement science is essential to understanding frontier AI and its national security implications.” The quote emphasizes that CAISI’s mandate is metric, not policy-driven — the goal is to objectively measure capability thresholds, not to dictate market access conditions.

The agreements are structured flexibly — they include clauses allowing rapid response to future AI advancements without requiring renegotiation. Test results remain in the classified channel, but NIST collaborates with labs on voluntary product improvements and sharing of information on international competitiveness.

Why is this a turning point for frontier AI regulation?

The consolidation of all five leading US frontier labs (Anthropic, OpenAI, Google DeepMind, Microsoft, xAI) under a single government evaluation framework is a structural change. Until 18 months ago, government evaluations of AI models were ad-hoc and based on voluntary disclosure. CAISI is now becoming the de facto national laboratory for frontier AI assessment.

Practical industry implications: labs must now maintain classified testing pipelines, document capability claims in a manner verifiable through CAISI methodology, and expect pre-release government review for extreme capability bumps. For the EU AI Office and UK AI Safety Institute, this is a reference model — a formal pre-deployment testing obligation with a “removed safeguards” testing mechanism that EU AI Act Article 51 (general-purpose models with systemic risk) has not yet operationalized at this level of detail.