Google DeepMind, Microsoft, and xAI have agreed to give the U.S. government early access to frontier AI models for national security testing before those systems are widely released.
The agreements expand the role of the Commerce Department's Center for AI Standards and Innovation, or CAISI, which sits inside NIST. CAISI says the work will include pre-deployment evaluations, post-deployment assessments, and targeted research into advanced model capabilities.
NIST's announcement says the partnerships cover Google DeepMind, Microsoft, and xAI, and that evaluations can include testing in classified environments. Reuters reports that the goal is to check new models for national security risks before public release.
The new checkpoint
This is not a consumer product review. CAISI is being positioned as the U.S. government's main technical interface with frontier AI labs. The agency says it has already completed more than 40 evaluations, including reviews of unreleased state-of-the-art systems.
The important detail is access. According to NIST, developers frequently provide models with reduced or removed safeguards for national security-related testing. That gives evaluators a clearer view of what a model can do when its normal product restrictions are not in the way.
That kind of access is useful, but it also raises the stakes. A lab that shares a rawer model with government testers is trusting that the process, the facility, and the feedback loop are secure enough for extremely sensitive systems.
Why labs are saying yes
For AI companies, voluntary review may be the least disruptive version of oversight. It lets companies keep shipping, while giving Washington a way to claim it is measuring risks before the public gets access.
The timing also matters. Frontier models are increasingly discussed as cybersecurity tools, military-support systems, and economic infrastructure. Once models can meaningfully help with vulnerability discovery, malware analysis, chemical or biological research, or autonomous planning, governments will not treat releases like ordinary software launches.
Google, Microsoft, and xAI joining the process also makes CAISI harder to ignore. OpenAI and Anthropic already had earlier arrangements with the government. With more major labs participating, pre-release testing starts to look less like a one-off safety pledge and more like emerging industry plumbing.
The policy tradeoff
The upside is straightforward: independent testing can catch dangerous capabilities before a model becomes a public product. It can also give policymakers better evidence than press demos, benchmark claims, or company assurances.
The risk is that voluntary access becomes informal regulation without clear public rules. If government testing shapes which models launch, when they launch, or what safeguards are required, companies and users will need more transparency about the standards being applied.
There is also a competition angle. NIST says the work will improve the government's understanding of international AI competition. That means model review is not only about safety. It is also about national advantage, defense readiness, and which companies become trusted AI suppliers to the state.
What to watch next
The practical question is whether CAISI feedback changes releases in visible ways. Do models get delayed? Do labs add new refusal behavior? Do some capabilities move into restricted enterprise or government-only channels? Those signs will show whether the checkpoint has teeth.
The other question is whether Congress or the White House turns this voluntary process into a more formal requirement. For now, CAISI gets deeper access and the labs get a cooperative path through Washington. That bargain may hold only as long as the next frontier model does not trigger a public safety or security incident.
