In the space of four days, the U.S. government announced two parallel sets of deals with leading AI companies, which together define the two directions Washington wants to take simultaneously – to test AI for national security risks before the public sees it and to deploy AI directly into the military’s most secretive networks. The Center for AI Standards and Innovation — CAISI , the entity under the Commerce Department’s National Institute of Standards and Technology that inherited the mission of the former AI Security Institute — announced new deals with Google DeepMind, Microsoft and Elon Musk’s xAI
See also: Google DeepMind: Employees against military AI contracts
These are based on renewed agreements with Anthropic and OpenAI dating back to 2024, updated to reflect guidance from Commerce Secretary Howard Lutnick and the AI Action Plan . Under the CAISI agreements, the three companies will deliver their cutting-edge AI models to government evaluators before they are released publicly.
The assessments examine capabilities and risks related to national security. To conduct a full assessment, developers often provide CAISI with models that have reduced or removed security safeguards — a design choice that allows evaluators to examine what a model can do at its maximum capability, not what it will do under commercial security controls.
Evaluators from across the federal government are participating, coordinated through the TRAINS Working Group formed by CAISI, an interagency body focused specifically on AI national security concerns. CAISI said it has completed more than 40 such assessments to date. The agreements explicitly support testing in classified environments and were written with the flexibility to adapt quickly as AI capabilities continue to advance.
“Independent, rigorous measurement science is essential to understanding cutting-edge AI and its implications for national security,” said CAISI Director Chris Fall. Fall was appointed to lead CAISI following the reported removal of Collin Burns — a former Anthropic researcher — from the director role after just four days.
See also: Anthropic Mythos pushes White House to consider pre-publications for high-risk AI models
The personnel transition at the top of CAISI reflects a broader institutional shift. Under the Biden administration, the AI Security Institute focused on security standards, definitions, and voluntary safeguards. Under Trump, CAISI has shifted its emphasis to accelerating AI and assessing national security capabilities. The essence of what the evaluators do — reviewing robust models before release — has not changed.
The context for why they’re doing it is unclear. The latest announcement comes four days after the Department of War (formerly the Department of Defense) announced agreements with eight leading AI companies to deploy their models directly to the military’s classified networks for operational use. The companies approved include SpaceX, OpenAI, Google, NVIDIA, Reflection, Microsoft, Amazon Web Services , and Oracle.
The networks involved are classified as Impact Level 6, which covers classified data, and Impact Level 7, which refers to the most strictly restricted national security systems. The stated goals are data synthesis, situational awareness enhancement, and warfighter decision support. The War Department’s announcement is conspicuously absent from the coverage of what it actually means.
Anthropic is not on the list. The company that first deployed AI models into classified Pentagon systems — through a Palantir under the Maven Smart System — is being excluded after a disagreement over safeguards governing the military and surveillance use of its AI. The Pentagon had previously labeled Anthropic a “supply chain risk,” a designation typically given to foreign entities that pose national security concerns.
See also: Pentagon: Agreements with 7 technology companies for AI in secret systems
A March 2026 federal order reversed that designation, but did not restore Anthropic's position as the Pentagon's AI supplier.
