Government Will Pre-Test AI from Google, Microsoft, xAI

The federal government just got a key to the AI factory

On May 5, 2026, the Center for AI Standards and Innovation — CAISI, sitting inside the Commerce Department — announced agreements with Google DeepMind, Microsoft, and xAI to evaluate their frontier AI models before those models ship to the public. The deals expand on prior 2024 partnerships with OpenAI and Anthropic and pull the four largest US-based frontier developers under the same testing umbrella.

If you run a small business, the headline reads like Beltway noise. It isn’t. The thing CAISI is doing — looking inside an AI model before the public uses it — is the closest thing the AI industry has to a UL listing or a USDA stamp. For SMBs picking AI tools, who tested the model and how thoroughly is becoming a real buying signal.

What CAISI actually agreed to

Three things are worth knowing about the May 5 deals.

Pre-deployment access. The agreements give CAISI the ability to evaluate models before public launch — including state-of-the-art systems that haven’t been announced yet. Per the NIST announcement, the center has already completed more than 40 model evaluations under prior frameworks.

Models with safeguards stripped. This is the unusual part. To stress-test what a model can actually do under adversarial pressure, developers hand CAISI versions with reduced or removed safeguards. The point is to find national-security-relevant capabilities the production model would refuse to perform — biosecurity, cyberweapons, chemical synthesis — and figure out where the floor really is.

Classified-environment testing. Some of this work happens in classified facilities, with results shared back to the developer through information-sharing channels designed to drive voluntary product changes. CAISI Director Chris Fall framed the rationale plainly in the NIST release: “Independent, rigorous measurement science is essential to understanding frontier AI and its national security implications.”

Microsoft and xAI both posted their own statements. Microsoft’s Chief Responsible AI Officer Natasha Crampton wrote that “ongoing, rigorous testing is essential to building trust and confidence in advanced AI systems,” and noted the company is also working with the UK AI Security Institute on parallel evaluation methods.

Why this is a real story, not theater

There are two ways to read voluntary government testing of corporate AI: as a meaningful safety check, or as a rubber stamp where companies pick what to share. The honest answer for May 2026 is “both, and we won’t know which dominates for another year.” But several things make this more substantive than the typical industry-government photo op.

The testing is structured around models with safeguards removed, which means CAISI is looking at raw capability rather than the polite version. The center publishes evaluation findings with developers and shares interagency feedback through a TRAINS Taskforce that pulls in expertise across DoD, DHS, and other national-security shops. And the agreements explicitly cover models that haven’t been launched, not just yesterday’s ChatGPT.

That last piece matters. Until 2024, federal evaluation of AI models meant the government bought GPT-4 like everyone else and poked at it after the fact. The new model is closer to how FDA looks at a drug before it hits the pharmacy — not perfect, not the same level of legal authority, but a meaningful shift in posture.

The flip side is what CAISI explicitly does not do. It is not a regulator. It does not block model releases. There is no public certification or seal a developer can put on a product. If CAISI finds something concerning, the response runs through “voluntary product improvements” — meaning the developer chooses what to fix and when. The Trump administration’s AI Action Plan is consistent on this: light-touch federal posture, with measurement and standards rather than mandates.

What this means if you buy AI for a small business

For an SMB owner staring at four AI vendor pitches and trying to figure out which to trust, the CAISI news is useful in two specific ways.

You now have a tier of “tested” frontier developers. OpenAI, Anthropic, Google DeepMind, Microsoft, and xAI are all under formal CAISI evaluation arrangements. That doesn’t make their products risk-free — it does mean their underlying models have been examined by people whose job is to find catastrophic failure modes. A reseller building a chatbot on top of GPT-4 or Claude is shipping a product whose foundation has been independently tested. A reseller building on a model nobody has ever evaluated has not.

The “vendor trust” question is now answerable with a sentence. When you’re vetting an AI tool — for an AI Employee, an intake widget, a back-office automation — ask the seller two questions: which underlying model powers it, and is that model from a developer with a CAISI evaluation in place. If the answer is “we don’t know” or “a custom open model,” that is fine, but it is a different risk profile than “Anthropic Claude 4.6 with a system prompt.” Both can be valid choices. They are not the same choice.

This was harder to evaluate even six months ago. There were too many models, no shared testing baseline, and the marketing was indistinguishable. The CAISI list is short, durable, and tells you something real.

What is missing from the conversation

A few things deserve to be said out loud.

The agreements don’t cover open-source frontier models. DeepSeek, Kimi K2, Mistral, Meta’s Llama family — all the open weights that have been reshaping the cost floor — are not under any CAISI arrangement. That is not a knock on those models. It is a fact about the boundary of what CAISI does. SMBs running on open-weight models are operating in a space that the federal evaluation regime does not currently cover.

CAISI’s testing focuses on national-security risks. That is not the same as testing for the things SMBs actually care about: hallucination rates on real tasks, prompt injection resistance, data leakage from RAG pipelines, refusal behavior on legitimate queries. CAISI evaluations help with the catastrophic-tail-risk question. They do not tell you whether a model will give your customer the wrong dispatch address.

And the “voluntary” part of these deals is doing a lot of work. There is no enforcement, no sanction, no public scoring. It is a measurement regime in a country that has — for now — chosen not to regulate frontier models directly.

What to actually do this week

If you are running a small business and you saw the CAISI headline scroll past, the practical to-do list is short.

Inventory your AI dependencies. Make a list of the AI tools you pay for and the underlying model each one uses. Most vendors will tell you in their docs or sales conversations — if they refuse, that is a signal. We covered the broader AI agent security risk picture in a recent post.

Treat “tested model” as one input, not the whole answer. A CAISI-evaluated model is a stronger foundation than an unevaluated one for high-stakes use cases. It is not a guarantee, and it does not override the things you should still test yourself: accuracy on your real data, response quality on your customer cases, and whether the vendor can explain what happens when the model fails.

Watch what gets published. CAISI is supposed to share findings publicly when it can. The first round of post-evaluation reports under these new agreements will tell us a lot about how serious the regime really is. If the published findings are detailed and uncomfortable for the vendors, the program is doing real work. If they are polished and vague, you have your answer there too.

The federal government testing AI before launch is, on balance, good for small business buyers. It puts a floor under the foundation models powering the tools you depend on. It is not a substitute for paying attention. If you want help thinking through which AI tools actually fit your shop, get in touch — vetting AI vendors is most of what we do these days.