Microsoft, Google, xAI agree to share AI models with White House for security reviews – Image for illustrative purposes only (Image credits: Pixabay)
Washington — Cutting-edge artificial intelligence systems now face closer scrutiny from federal evaluators before reaching the public. Google DeepMind, Microsoft, and xAI reached agreements with the Department of Commerce’s Center for AI Standards and Innovation on May 5 to share early versions of their models. Officials aim to probe these tools for national security risks through targeted testing and research.[1][2]
New Pacts Enable Deeper Government Access
The deals mark an expansion of prior industry partnerships, renegotiated to align with directives from Commerce Secretary Howard Lutnick and the America’s AI Action Plan. Developers will hand over models ahead of public release, often with safeguards dialed back or removed. This setup allows testers to explore national security-related capabilities without typical restrictions.
CAISI, housed within the National Institute of Standards and Technology, leads these efforts. The center positions itself as the government’s main liaison for AI testing and research. Evaluations occur in controlled settings, including classified ones when necessary, to match the pace of rapid AI progress.[1]
CAISI’s Growing Role in AI Oversight
Established to foster standards and innovation, CAISI has already completed more than 40 evaluations on state-of-the-art models, some still unreleased. It coordinates with agencies like the Defense Department and intelligence community through the TRAINS Taskforce, a group of interagency experts focused on AI risks. These collaborations drive voluntary improvements in products and sharpen government insights into global AI competition.
“Independent, rigorous measurement science is essential to understanding frontier AI and its national security implications,” CAISI Director Chris Fall said. “These expanded industry collaborations help us scale our work in the public interest at a critical moment.”[1][3] The center also assesses vulnerabilities from adversarial systems and supports international positioning for U.S. standards.
Microsoft highlighted its commitment in a statement, noting plans to partner with government scientists on tests that “probe unexpected behaviors.” The company intends to build shared datasets and workflows for ongoing assessments. Such steps echo Microsoft’s recent pact with the UK’s AI Security Institute.[2]
Drivers Behind the Security Push
Washington’s moves respond to escalating worries about AI’s dual-use potential. Recent releases, like Anthropic’s Mythos, have spotlighted hacking enhancements that could empower cybercriminals or disrupt critical infrastructure. Officials seek to spot threats from cyberattacks to military applications before widespread deployment.
These agreements fulfill pledges from the Trump administration’s July 2025 AI Action Plan. They build on 2024 arrangements with OpenAI and Anthropic, when CAISI operated under a different name. The timing aligns with discussions of broader oversight, possibly via executive order, to convene tech leaders and policymakers.[4]
Precedents and Future Horizons
Past evaluations have informed best practices without mandating changes, emphasizing voluntary cooperation. CAISI’s work extends to monitoring deployed systems and countering foreign influences in AI. Recent activities include partnerships with groups like OpenMined for secure testing methods.[5]
As AI evolves, these pacts offer a framework for balanced advancement. They equip policymakers with data to navigate risks while preserving innovation’s edge. Whether this model scales to more firms or formalizes further remains a key watchpoint in the coming months.
Key Elements of the Agreements
- Pre- and post-deployment model evaluations
- Access to versions with reduced safeguards
- Focus on national security capabilities and risks
- Interagency input via TRAINS Taskforce
- Over 40 prior assessments completed
