AI Model Security Reviews: Microsoft, Google, xAI to Share
AI Model Security Reviews give U.S. officials early access for predeployment testing, raising scrutiny that may slow releases and reshape sector flows.
KEY TAKEAWAYS
- CAISI secured early access to frontier models from Microsoft, Google DeepMind and xAI for security reviews.
- Officials will run pre-deployment testing on versions with reduced safeguards to probe national-security risks.
- Program extends prior voluntary reviews and adds scrutiny that could complicate model releases.
HIGH POTENTIAL TRADES SENT DIRECTLY TO YOUR INBOX
Add your email to receive our free daily newsletter. No spam, unsubscribe anytime.
Microsoft, Google DeepMind and xAI agreed on May 5, 2026 to give the U.S. government early access to new AI models for AI model security reviews and pre-deployment safety evaluations. This access allows officials to test national-security capabilities on versions with reduced safeguards.
Predeployment Testing and Regulatory Context
The Center for AI Standards and Innovation (CAISI), part of the Commerce Department’s National Institute of Standards and Technology (NIST), signed agreements with Microsoft, Google DeepMind and xAI to obtain early, pre-public-release access to frontier AI models. CAISI will conduct safety evaluations and targeted research on national-security risks.
These evaluations will focus on model capabilities, security vulnerabilities and national-security implications. Developers will provide versions with safety guardrails reduced or removed so the center can probe potential risks.
This program extends a voluntary federal approach that began in 2024, when the predecessor U.S. AI Safety Institute reached similar agreements with OpenAI and Anthropic. CAISI has completed more than 40 evaluations, including work on unreleased models, demonstrating its ongoing role in assessing advanced AI systems.
The arrangements build on a Biden administration executive order requiring developers to submit internal safety evaluations for advanced models. Separate reporting on May 4 indicated the Trump administration considered a Pentagon-led AI safety framework and a possible executive order to allow federal access to commercial models, but the White House described those reports as speculation.
Tech-policy experts have warned the process could create deployment hurdles, while national-security officials have highlighted cyber and misuse risks. These dynamics may increase scrutiny of model releases and how government agencies gain access to them.
