Google Microsoft and xAI agree to let US test AI models before release
Google DeepMind, Microsoft, and Elon Musk’s xAI have signed agreements with the U.S. Commerce Department to give the government early access to their frontier AI models for national security testing before public deployment.
The Center for AI Standards and Innovation, or CAISI, announced the deals on Tuesday. The agency will conduct pre-deployment evaluations and targeted research to assess model capabilities and security risks. Microsoft confirmed it will work with government scientists to test systems “in ways that probe unexpected behaviors,” including developing shared datasets and testing workflows.
This builds on CAISI’s earlier agreements with OpenAI and Anthropic from 2024, which have been renegotiated to align with Commerce Secretary Howard Lutnick’s directives and the administration’s AI Action Plan.
The timing is not accidental. Government concern has been escalating since Anthropic unveiled Claude Mythos Preview last month, a model specifically designed to find and exploit software vulnerabilities. Mythos is powerful enough that Anthropic restricted its rollout to a select group of companies under a cybersecurity initiative called Project Glasswing. Anthropic CEO Dario Amodei was called to the White House to discuss the model days after launch, even though the Defense Department had previously flagged Anthropic as a supply chain risk.
The CAISI agreements are voluntary. There is no legal requirement for AI companies to submit their models for government review before release. The Biden-era executive order on AI safety was revoked by Trump in early 2025, and no binding federal legislation has replaced it. What exists now is a patchwork of voluntary commitments and agency-level agreements.
Beyond CAISI, the White House is weighing a new AI working group that would formalize oversight procedures and bring together tech executives and government officials. The New York Times first reported those discussions, and CNBC confirmed the details with a source close to the talks. The group could be established through executive order, though the White House told CNBC any policy announcement would come directly from Trump.
Notably absent from the agreement: Anthropic itself, despite being a CAISI partner since 2024. The company that triggered the current wave of government alarm is already operating under a renegotiated deal from the earlier round. Its Mythos model remains in limited release.
Microsoft also signed a separate agreement with the UK’s AI Security Institute, suggesting the company is building pre-deployment review into its standard release process across jurisdictions.
The fundamental tension here is clear. The U.S. government is building oversight infrastructure through voluntary cooperation, not regulation. That works when companies want to cooperate. It stops working the moment a company decides to release first and deal with the consequences later. The entire framework relies on goodwill from the same companies racing to build systems that worry officials in the first place.