Google DeepMind, Microsoft and Elon Musk's xAI have agreed to give the US government pre-release access to new frontier AI models for security evaluation, the Commerce Department's Center for AI Standards and Innovation said Tuesday. CAISI will run what it calls "pre-deployment evaluations and targeted research" on the models before they ship to the public. The center has already conducted 40 reviews since it began evaluating models from OpenAI and Anthropic in 2024. The expansion roughly doubles the number of frontier labs feeding unreleased systems into a federal review pipeline.
The May 5, 2026 announcement also confirms that OpenAI and Anthropic have renegotiated their existing CAISI partnerships, according to Bloomberg, to align with priorities in President Donald Trump's AI Action Plan. That detail matters: the prior arrangements were struck under the Biden administration's voluntary safety commitments, and the new terms reset the relationship around the current White House's framework. CAISI itself is the rebranded successor to the AI Safety Institute, with "standards" replacing "safety" in the name.
"Independent, rigorous measurement science is essential to understanding frontier AI and its national security implications," CAISI director Chris Fall said in the press release. "These expanded industry collaborations help us scale our work in the public interest at a critical moment." The phrasing — measurement science, national security — signals where the program's center of gravity has moved.
Key facts
- 01Google DeepMind, Microsoft and xAI will give CAISI pre-release access to new frontier models for security evaluation.
- 02CAISI, housed in the Commerce Department, has conducted 40 reviews since beginning evaluations of OpenAI and Anthropic models in 2024.
- 03OpenAI and Anthropic renegotiated their existing CAISI partnerships to align with Trump's AI Action Plan.
- 04The arrangement was announced Tuesday, May 5, 2026, by the Center for AI Standards and Innovation.
- 05Trump is weighing an executive order to bring tech executives and federal officials together to oversee new AI models.
Google DeepMind, Microsoft and xAI did not previously have formal pre-release review agreements with CAISI's predecessor on the same footing as OpenAI and Anthropic. Bringing all five major US frontier labs into one evaluation regime gives the federal government a consistent vantage point on capability jumps, jailbreak resistance, and dual-use risk across the leading models. It also gives the labs political cover: every new release now ships with a federal security check stamped on it.
“CAISI has run 40 reviews of frontier models since 2024, and is now adding Google DeepMind, Microsoft and xAI to a program that previously covered only OpenAI and Anthropic.”— Jaeden Schafer
The arrangement is voluntary, which is the load-bearing word. Companies can in principle withdraw, and CAISI's findings are not binding on a launch decision. But the optics of pulling out of a national-security review program are bad enough that the voluntary frame functions as soft regulation, particularly for labs already pursuing classified federal contracts.
The White House may push further. The New York Times reported Monday that Trump is considering an executive order that would "bring together tech executives and government officials" to oversee new AI models. That would move the relationship from lab-by-lab partnerships to a more formal coordinating body, potentially with a standing role for industry CEOs alongside federal officials.
For xAI, the deal is notable on its own terms. Musk has spent the past several weeks on the witness stand in the OpenAI nonprofit-conversion trial in Oakland, where xAI's competitive posture toward OpenAI has been a recurring theme. Signing onto the same federal review program as OpenAI puts xAI inside the same regulatory tent as the company Musk is suing.
For Google DeepMind and Microsoft, the move tracks with broader federal entanglement. Microsoft was among the vendors awarded classified Pentagon AI contracts last week, alongside Nvidia and AWS. Pre-release CAISI review and classified deployment work are different programs, but they point in the same direction: the largest US AI labs are increasingly operating as quasi-defense contractors, with the federal government as a structural customer and reviewer.
Skeptics will note what the program does not include. CAISI evaluations are not safety certifications, and the center has not published the methodology, scope or pass-fail criteria of its 40 reviews to date. There is no public registry of which models were reviewed, which findings were flagged, or whether any launch was delayed as a result. "Renegotiated to align with the AI Action Plan" also leaves room for the new terms to be narrower than the old ones, not broader. Until CAISI publishes more, the program's substance is taken on faith.
The expansion still resets the default. Six months ago, US frontier-model oversight was a patchwork of voluntary commitments, state laws and export controls. With Google DeepMind, Microsoft, xAI, OpenAI and Anthropic now all routing pre-release models through the same Commerce Department office, Washington has quietly built the closest thing the US has to a national AI regulator — and it did so without passing a law.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
