New agreements allow the U.S. government to test advanced AI models for national security risks before public deployment.
The Center for AI Standards and Innovation (CAISI) at the Department of Commerce’s National Institute of Standards and Technology (NIST) announced agreements with Google DeepMind, Microsoft, and xAI to provide government access to their AI models for testing national security risks before they are released to the public.
The participating companies will submit selected frontier AI models to CAISI for evaluation before release. Evaluations to identify risks related to cybersecurity, misuse, and model capabilities will occur before and after deployment.
The agreements allow CAISI to test models under conditions that may include the reduction or removal of safety safeguards. This enables government researchers to assess how models behave without built-in restrictions and to identify potential vulnerabilities that could be exploited.
CAISI said it previously conducted more than 40 evaluations of AI models, including systems that were not yet publicly available at the time of testing. The new agreements formalize and expand that process by establishing ongoing collaboration with specific companies.
Under the agreements, companies may also provide technical information about their models and work with CAISI on research related to model risks. The announcement does not specify which models qualify for testing, what thresholds determine inclusion, or whether all future models from participating companies must be submitted.
The agreements aren’t law. They do not compel participation or express penalties for non-compliance. Further, the testing agencies don’t have the authority to delay or block releases.
CAISI operates within the U.S. Department of Commerce and is responsible for evaluating advanced AI systems and developing measurement standards. The program focuses on identifying risks to national security, including the potential for models to assist in cyberattacks or to generate harmful content.
The announcement follows reports that the White House is considering a policy requiring pre-release testing of AI models.
