On May 5, 2026, the Center for AI Standards and Innovation — CAISI, the Commerce Department body inside NIST — announced pre-deployment evaluation agreements with Google DeepMind, Microsoft, and xAI. The agreements expand on the 2024 pacts CAISI's predecessor signed with OpenAI and Anthropic. As of this week, five of the largest U.S. frontier labs have formal agreements giving the federal government access to evaluate models before public release.
The structure of the deal is the part to read carefully. Labs provide models to CAISI for "pre-deployment evaluations and targeted research to better assess the frontier AI capabilities and advance the state of AI security." More importantly, labs frequently provide CAISI with models that have "reduced or removed safeguards" — meaning CAISI is testing what the model can do absent the public-release guardrails.
The testing is not theoretical. CAISI has completed more than 40 frontier-AI evaluations to date. The agreements support testing in classified environments, and were drafted with what NIST calls "the flexibility required to rapidly respond to continued AI advancements."
Why this matters: the U.S. government is now structurally upstream of every major frontier model launch. The agreements are voluntary. They are also the new default. The next frontier lab that ships a flagship model without a CAISI eval is going to look like a regulatory outlier, not an industry standard.
What to watch: whether Meta, Mistral, and the smaller-but-frontier U.S. labs sign equivalent agreements in the next 90 days. Also worth tracking — whether the evaluations begin influencing release timing, capability claims, or safety disclosures in public-facing model cards. The first model that delays a release by a quarter to address a CAISI finding is going to be the moment this program becomes regulation in everything but name.
