Google, Microsoft, xAI Frontier Models Must Be Tested by US Government Before Release

The Artificial Intelligence Standards and Innovation Center (CAISI) under the U.S. Department of Commerce ( announced) Tuesday, has signed new agreements with Google DeepMind, Microsoft, and xAI, allowing the federal government to conduct national security testing before the public release of cutting-edge AI models.

This arrangement moves government assessments ahead of model deployment. CAISI stated that the new agreements will support pre-release assessments, post-release evaluations, and targeted research to determine the capabilities and potential risks of cutting-edge models.

CAISI had previously announced similar collaborations with Anthropic and OpenAI in August 2024 and indicated that it has completed over 40 assessments, including advanced models not yet publicly released. With the addition of Google DeepMind, Microsoft, and xAI, the release process of cutting-edge AI companies is entering the government testing framework more systematically.

Models participating in testing may not necessarily be the final commercial versions for end-users. CAISI indicates that to fully assess capabilities and risks related to national security, developers often provide models with reduced or removed safety protections.

This allows the government to observe behaviors closer to the model's extreme states and also integrates AI companies' safety designs, capability disclosures, and release tempos into the policy framework earlier. For top developers like Google, Microsoft, and xAI, the competition in cutting-edge models is not just about the competition of computing power, data, and product experience but also begins to be influenced by government testing processes.

According to the announcement, under the leadership of Secretary Howard Lutnick, CAISI has been designated as the primary point of contact within the U.S. government for testing, collaborative research, and best practice development for commercial AI systems. One day before the announcement, there were reports that the White House is considering new AI model guardrails and review processes, including the possibility of establishing a working group composed of industry executives and government officials through an executive order.