Trump administration weighs Pentagon-led safety tests for AI models used by governments

The Trump administration is weighing a framework that would require advanced artificial intelligence models to undergo safety testing before being deployed across federal, state and local governments, according to reports and people familiar with the discussions.

The talks have taken place within the White House’s Office of the National Cyber Director (ONCD) amid growing concern about security risks posed by newer AI systems, including Anthropic’s Mythos Preview.

Pentagon role under discussion

One option under review would have the Pentagon lead safety assessments of AI models intended for public-sector use. Supporters say the goal would be to add an additional layer of government oversight to identify and mitigate potential security vulnerabilities before agencies adopt the technology.

According to the report, the ONCD held two meetings last week—one with technology and cybersecurity companies and another with trade groups—to discuss broader security implications raised by advanced AI models.

Policy tensions and echoes of the prior AI order

Although the administration previously revoked the Biden administration’s AI executive order, some of the ideas under discussion resemble parts of that earlier approach, which required AI developers to submit the results of internal safety and security testing.

Deliberations remain fluid, and potential proposals for executive action could change. Economic and tech policy stakeholders have raised concerns that new requirements could complicate AI deployments, while national security officials have focused on the risk of AI-enabled cyberattacks.