U.S. model-testing deals bring major frontier labs under CAISI review

The U.S. government's frontier-model testing program is becoming a meaningful launch-process story. NIST's Center for AI Standards and Innovation signed agreements in May with Google DeepMind, Microsoft, and xAI for national security testing, building on prior OpenAI and Anthropic relationships from 2024. Tom's Hardware and policy groups summarized the practical implication: every major U.S. frontier AI lab is now participating in voluntary pre-release or national-security-oriented model evaluation arrangements. This is not the same as EU-style binding product regulation, but it creates a route for Washington to inspect powerful systems before public deployment and before federal procurement decisions. The watch item is how much access CAISI gets to scaffolding, tools, and unreleased capabilities, because modern agent risk often comes from a model plus environment rather than a base model alone.

Key details: CAISI, NIST, May 5, 2026, Google DeepMind, Microsoft, xAI, OpenAI, Anthropic.

Continue swiping for more AI Brief stories.