Microsoft, Google, and xAI have signed agreements with the U.S. Commerce Department’s Center for AI Standards and Innovation, or CAISI, committing to submit new models for government safety evaluation before public deployment. Reuters said the move comes as Washington pays closer attention to the risks tied to advanced AI systems, especially after concern around Anthropic’s unreleased Mythos model.
CAISI keeps the testing role after its reorganization
CAISI was previously known as the U.S. AI Safety Institute, or AISI, during the Biden administration. It was led by technology adviser Elizabeth Kelly and focused on AI testing protocols and voluntary safety standards. After the Trump administration took office in 2025, the agency was reorganized and renamed CAISI, but its main mission stayed in place: evaluating frontier models and identifying security risks.
According to the report, CAISI has completed more than 40 evaluations so far, including assessments of several models that had not yet been released publicly. Microsoft said in a statement that it will work with U.S. government scientists to test AI systems by probing for unexpected behavior. The two sides will also build shared datasets and workflows for testing. Google and xAI did not respond to requests for comment.
Anthropic triggered concern but is not part of this deal
The immediate backdrop is Mythos, an unreleased Anthropic model that reportedly can identify zero-day vulnerabilities across major operating systems and browsers. That claim drew broad attention and raised alarms inside Washington over what highly capable models might do if released without checks. Even so, Anthropic was not included in this round of agreements, which only covered Microsoft, Google, and xAI.
The report said companies that stay outside such arrangements could face pressure if the U.S. government tightens its review framework. At the same time, Washington is moving on two tracks. CAISI is focused on technical governance and pre-deployment review, while the Defense Department is pushing direct adoption and procurement of AI capabilities for military systems.
Pentagon procurement push is moving in parallel
Last week, the Pentagon announced agreements with seven AI companies to allow AI capabilities on classified Defense Department networks. The list included AWS, Google, Microsoft, Nvidia, OpenAI, SpaceX, and Reflection. Hours later, Oracle joined, bringing the total to eight companies.
Anthropic was absent from that list as well. The report said one possible reason was the company’s earlier refusal to accept a clause allowing lawful access without use restrictions, which may have contributed to it being treated by the U.S. government as a supply-chain risk. With model testing standards advancing at the same time as military AI procurement, willingness to submit systems for review is becoming a competitive factor for developers.