AI oversight

US to test Google, Microsoft and xAI models before release

Voluntary Commerce Department agreements give CAISI early access to frontier systems as Washington weighs security risks from more powerful AI

Source language: English
0
US to test Google, Microsoft and xAI models before release
Location
Washington
Washington, District of Columbia, United States
Google, Microsoft and xAI agreed to let a Commerce Department center evaluate new AI models before public release, expanding federal safety testing.
AI Safety Artificial intelligence Google Microsoft xAI

Google, Microsoft and xAI agreed to let a Commerce Department center evaluate new AI models before public release, expanding federal safety testing.

WASHINGTON — The US Department of Commerce will safety test new artificial intelligence models from Google, Microsoft and xAI before they are released or deployed, under voluntary agreements announced Tuesday by its Center for AI Standards and Innovation.

The deals give federal scientists access to frontier AI systems from three major developers at a time when Washington is trying to assess security risks from increasingly capable models without imposing broad mandatory rules on the sector.

CAISI said the work will include testing, collaborative research and development of best practices for commercial AI systems, with evaluations focused on capabilities and security. The center’s director, Chris Fall, said the agreements would expand government work “in the public interest at a critical moment.”

The agreements extend earlier arrangements reached under the Biden administration with OpenAI and Anthropic. The Commerce testing center, then known as the US Artificial Intelligence Safety Institute, has been used as a federal hub for AI evaluations and voluntary safety standards.

CAISI said Tuesday it has completed more than 40 previous evaluations, including tests of “state-of-the-art models that remain unreleased.” It did not identify those models. The agency has said developers often provide versions of models with some guardrails stripped back so evaluators can probe for national security risks.

Microsoft said it would work with US government scientists to test AI systems “in ways that probe unexpected behaviors” and to develop shared data sets and workflows for model evaluation. The company also said it has signed a similar agreement with the United Kingdom’s AI Security Institute. Google declined to comment to Al Jazeera, and xAI did not immediately respond to its request for comment.

The announcement comes amid growing concern in Washington about how advanced AI systems could be misused in cyberattacks or military settings. Recent attention has focused in part on Anthropic’s Mythos model and questions about whether powerful systems could supercharge hackers before governments and companies fully understand their risks.

The testing pacts also follow a Pentagon agreement with seven technology companies — including Google, Microsoft, Amazon Web Services, Nvidia, OpenAI, Reflection and SpaceX — to use AI systems across classified computer networks. The Defense Department said that arrangement is intended to help “augment warfighter decision-making in complex operational environments.”

The move marks a notable expansion of federal access to company models under President Donald Trump, whose administration has otherwise emphasized reducing regulatory barriers to AI development. Trump last year signed executive orders forming the basis of an AI Action Plan that he said would “remove red tape and onerous regulation” and help the United States win through AI advancement and control.

The public announcement did not name the specific Google, Microsoft or xAI models that will be reviewed, nor did it specify how much of CAISI’s findings will be released. For now, the agreements put more of the most powerful commercial AI systems inside a federal testing process before they reach wider use.

More from this section

Tech news

Figures mentioned

More from this location

Related tags

Related articles

Shared tag: AI Safety AI trust
Friendlier AI chatbots may be less accurate, study suggests

Oxford Internet Institute researchers found models tuned to sound warmer made more mistakes and were more likely to affirm false beliefs

Apr 29, 2026
Shared tag: Artificial intelligence
Google plans up to $40 billion Anthropic investment

The agreement starts with $10 billion and could add $30 billion more if performance milestones are met, deepening a partnership around Claude and cloud infrastructure

Apr 28, 2026 anthropic features
Shared tag: Artificial intelligence U.S.-China tech race
China Was Denied Access to Anthropic’s Newest AI, Report Says

The limited source summary says newer Anthropic and OpenAI models are widening the U.S. edge over China, but key details remain unavailable

May 12, 2026 anthropic features
Shared tag: Artificial intelligence Artificial intelligence
AI agents are starting to shop for consumers. Experts urge caution

Retailers and payment companies are testing agentic commerce, but specialists warn that autonomous purchases can expose shoppers to costly errors and data risks

Apr 28, 2026 United States
Shared tag: Artificial intelligence Artificial intelligence
Microsoft CEO Questioned in Court Over OpenAI Control

The court exchange centered on whether Microsoft had too much influence over OpenAI, an organization described as intended to operate as an A.I. nonprofit

May 12, 2026
Shared tag: Artificial intelligence Artificial intelligence
Google workers urge Pichai to reject classified Pentagon AI work

In an open letter, employees working on AI systems warned that a Defense Department deal could put Google technology into harmful military uses

Apr 28, 2026

Comments (0)

Please log in to comment.
No comments yet.