CAISI Pre-Release AI Testing

Voluntary federal pre-release testing of frontier AI is becoming a procurement filter. Manufacturers in regulated supply chains need to know which models clear it.

Executive Summary

The Center for AI Standards and Innovation (CAISI), successor to the AI Safety Institute, signed agreements with Google DeepMind, Microsoft and xAI to test frontier models before public release.
Anthropic was not part of this round, echoing its recent exclusion from a Pentagon contract and reinforcing a pattern of tier-one access governance.
For brand manufacturers in food, pharma and automotive, CAISI status is shaping up as a de facto AI procurement signal even though the program is voluntary.

On May 5 the Trump administration confirmed that CAISI, the rebranded successor to the federal AI Safety Institute, had signed pre-deployment evaluation agreements with three frontier AI labs: Google DeepMind, Microsoft and xAI. The reviews cover cybersecurity exposure, biological and chemical uplift risk, and alignment failures, and they happen before models reach customers. Anthropic, notably, is absent from the announced cohort.

From Safety Institute To Standards And Innovation

The renaming is not cosmetic. The original AI Safety Institute under the previous administration framed its mission around catastrophic risk; CAISI repositions the same machinery around standards-setting and US industrial competitiveness. The testing scope has not shrunk, but the political packaging has shifted from precaution to capability assurance. That matters because standards bodies historically outlive the administrations that create them and end up embedded in federal procurement language.

Critics call this regulatory capture wrapped in patriotism: a small club of labs gets a government stamp that doubles as a moat. Supporters call it pragmatic, arguing that voluntary pre-release testing is the only realistic way to surface frontier risks without freezing US competitiveness against Chinese labs. Both readings can be true at once, and both produce the same downstream effect for buyers.

Why Manufacturers Should Care About A Voluntary Program

Voluntary today rarely stays voluntary. Federal contractors, especially in defense-adjacent supply chains, are already being asked which AI vendors they use. The next logical step is asking which of those vendors have completed CAISI evaluation. A brand manufacturer running demand forecasting on a frontier model that has not been reviewed will not be blocked, but it will increasingly have to explain that gap to risk and compliance teams, and to enterprise customers running their own AI vendor questionnaires.

This connects directly to the Mythos restricted-access pattern observed earlier in the same week, where high-capability AI features were gated to selected partners. The signal across both stories is consistent: top-tier AI access is moving from open self-service to vetted distribution, and the vetting is partly governmental.

Epinium data

Of the 60+ brand manufacturers we’ve onboarded in 5+ years, fewer than 1 in 5 has a documented AI vendor evaluation policy — meaning most are not equipped to track which models pass CAISI review or what that designation actually guarantees.

What A Procurement-Ready AI Policy Looks Like Now

Manufacturers should treat CAISI status the same way they treat ISO certifications on a supplier audit: not as proof of safety, but as a documented checkpoint. Useful next steps include mapping every active AI vendor to its evaluation status, asking vendors directly whether they have engaged with CAISI, and setting an internal review trigger for any model used in regulated workflows. None of this requires new headcount; it requires a one-page policy and a quarterly check.

What is CAISI and how is it different from the AI Safety Institute?

CAISI, the Center for AI Standards and Innovation, is the Trump administration’s renamed successor to the AI Safety Institute. It retains pre-deployment testing authority but reframes the mission around standards and US competitiveness rather than catastrophic risk reduction.

Are the CAISI agreements legally binding?

No. The agreements with Google DeepMind, Microsoft and xAI are voluntary memoranda. There is no statutory requirement to participate, and there are no public penalties for non-participation. Federal procurement preference is the more likely enforcement channel.

Why is Anthropic not in this round?

Anthropic was not announced as a signatory and was also recently excluded from a Pentagon AI contract. The pattern suggests friction between Anthropic and the current administration, though neither side has confirmed a formal break. Buyers should not over-read it, but should track it.

Does CAISI testing certify a model as safe to deploy?

No. The evaluations cover specific risk categories such as cyber, bio-chem uplift and alignment, but they do not certify suitability for any given commercial use case. Deploying brands still own the contextual risk assessment for their own workflows.

What should a brand manufacturer do this quarter?

List every AI vendor in active use, ask each one whether they have engaged with CAISI, and document the answer. Add a single line to your AI policy stating that frontier-model vendors in regulated workflows should disclose federal evaluation status. That is enough to be defensible.

The next twelve months will tell us whether CAISI becomes a serious gate or a press release. Either way, manufacturers who start tracking model evaluation status now will spend the back half of the year answering customer questionnaires from a position of evidence rather than improvisation. Sources: CNN, CNBC. See also our Epinium blog for the broader AI governance series.

Need an AI procurement framework that actually accounts for federal review? CAISI status is becoming a question your enterprise customers will ask, and most manufacturers cannot yet answer it without scrambling. See how Epinium Transform builds AI vendor governance into operations →