Google, Microsoft and xAI Join US Pre-Release AI Model Testing
· TexaHub · Regulation & Ethics, LLMs & Foundation Models, Geopolitics & Tech
New CAISI agreements bring all five leading US frontier labs into voluntary government testing, as the White House weighs whether to make pre-release review mandatory.
The agreements
On 5 May the Center for AI Standards and Innovation (CAISI), the AI evaluation unit within the Commerce Department's National Institute of Standards and Technology, announced security testing agreements with Google DeepMind, Microsoft and xAI [1][3]. According to MeriTalk, the agreements allow federal evaluation of models before public release, assessments after deployment, and joint research [1]. CSO Online reported that CAISI will conduct pre-deployment evaluations and targeted research and give feedback on potential safety improvements in collaboration with the UK AI Security Institute [2]. MT Newswires reported that CAISI receives unrestricted versions of the models to evaluate potential national security risks, and that the arrangements support testing in classified environments [5].
OpenAI and Anthropic already had agreements dating to 2024, and Tom's Hardware reported that both renegotiated them to align with priorities in the administration's AI Action Plan [3]. With the three new signatories, Tom's Hardware concluded, every major US frontier lab now takes part in voluntary pre-release government evaluation [3]. CAISI says it has completed more than 40 evaluations to date, including of unreleased state-of-the-art models [1].
"Independent, rigorous measurement science is essential to understanding frontier AI and its national security implications," said CAISI director Chris Fall, adding that the partnerships "help us scale our work in the public interest at a critical moment" [1].
Same function, new framing
The arrangement is not new in kind. In August 2024 the then US AI Safety Institute signed memoranda of understanding with Anthropic and OpenAI giving it access to major models before and after public release, with feedback coordinated with the UK institute [6]. Tom's Hardware noted that the center was set up in 2023 under the Biden administration as the AI Safety Institute and renamed last June, and that its core work, evaluating frontier models for cybersecurity, biosecurity and chemical weapons risks, has stayed largely the same [3].
The center's position remains fragile. It still lacks permanent legal standing, and draft legislation to codify it has not passed [3]. Fall took over after Collin Burns, a former Anthropic and OpenAI researcher, was pushed out four days into the job, according to Tom's Hardware [3]. The AI Action Plan, announced last July, directs CAISI to lead national security-related model assessments as part of an "AI evaluations ecosystem" [3].
A mandatory track under discussion
The voluntary deals arrived one day after reports that the White House was considering a mandatory pre-release review process by executive order [3]. Citing The New York Times, Tom's Hardware reported that the proposal would create an "AI working group" of executives and officials, that the NSA, the Office of the National Cyber Director and the Director of National Intelligence could oversee reviews, and that the system would give government early access without blocking release [4]. A White House official called talk of an order "speculation" [4]. CSO Online cited a Bloomberg report that the White House is preparing an order to set up a vetting system for all new AI models [2].
The reported catalyst is Anthropic's Mythos model, which the company has described as capable of finding thousands of critical software vulnerabilities [4]. Tom's Hardware also noted that some analysts question whether Mythos justifies that framing [4]. It is unclear how a mandatory regime would interact with the CAISI agreements [3].
A shift in direction
The administration revoked a Biden-era executive order on AI risks immediately after taking office [4]. Independent analyst Carmi Levy told CSO Online the announcements signal "a significant shift in policy direction from an administration that up until recently had been following a more laissez-faire approach" [2]. Dean Ball, a former White House AI adviser, told The New York Times that officials are trying to strike a "tricky balance" between avoiding overregulation and keeping pace with the technology [4]. Fritz Jean-Louis of Info-Tech Research Group told CSO Online that the shift enables "government-led testing of advanced models before and after deployment", which he said would strengthen visibility into autonomous behaviours [2].
Policy leadership is also changing. David Sacks, who led the deregulation push as AI czar, left the role in March, and Chief of Staff Susie Wiles and Treasury Secretary Scott Bessent have taken a more active role since [4]. Relations with Anthropic are strained in parallel: the Pentagon designated the company a supply chain risk in March, a move a federal judge later called "Orwellian" [3].
What it means for developers
For the labs, the new agreements cost little. They formalise access that the two largest model developers have granted since 2024, and they do not, as reported, give the government power to block a release. They do give Washington a consistent view across all five leading US labs, which will matter if a mandatory process follows. Access to unrestricted model versions is also the part of the arrangement most likely to draw scrutiny, because it is where the government sees capabilities that public releases are designed to suppress. The open questions are the ones the reporting leaves unresolved: whether an executive order goes ahead, which agency would lead reviews, and whether CAISI gets a statutory footing. For companies building on frontier APIs, a review step before release could mean longer gaps between a model's completion and its availability. The scale of that delay depends on details that have not yet been published.
Sources
[1] CAISI Enters Security Testing Agreements With Google, Microsoft, and xAI — MeriTalk, 5 May 2026
[2] US government agency to safety test frontier AI models before release — CSO Online, 6 May 2026
[3] Google, Microsoft, and xAI agree to let US government test AI models before public release — Tom's Hardware, 5 May 2026
[4] Trump administration considers mandatory pre-release vetting of AI models — Tom's Hardware, 4 May 2026
[5] Google DeepMind, Microsoft, xAI Sign AI Security Agreements With CAISI — MT Newswires, 5 May 2026
[6] U.S. AI Safety Institute Signs Agreements Regarding AI Safety Research, Testing and Evaluation With Anthropic and OpenAI — NIST, 29 Aug 2024