NSA spending billions to test AI models for national security risks

The NSA’s Artificial Intelligence Security Center has been testing frontier AI models to identify potential national security vulnerabilities, according to two sources familiar with classified intelligence estimates cited by The Washington Sun. The agency told lawmakers it is spending billions of dollars in taxpayer funds this year on these efforts, a figure described as significantly higher than previously known.

The revelation has prompted lawmakers to consider whether a more comprehensive federal AI regulatory system could cost the government tens of billions of dollars per year, the sources said. President Donald Trump has resisted calls for greater federal oversight of AI, but the NSA’s testing initiative has injected new urgency into the debate over the government’s role in reviewing models developed by some of the most powerful technology companies.

The NSA’s spending is driven in part by the cost of computing power and personnel. Purchasing the chips and infrastructure needed to run and test AI models has become increasingly expensive amid global chip shortages and surging demand. Anthropic has separately lined up computing deals that could cost as much as $517 billion, according to The Information, illustrating the scale of investment required.

Hiring top AI engineers has also proven costly, with some experts commanding compensation in the hundreds of millions of dollars—far outpacing what the government can offer. Despite the NSA’s reputation for technical expertise, the agency may struggle to compete with private-sector pay packages. Nathan Calvin, general counsel at AI advocacy organization Encode, described in-house AI evaluation as “extremely important and necessary” but acknowledged its high cost and the competitive hiring environment.

Some experts argue that AI developers should help fund independent government audits. Nat Purser, director of U.S. Policy at the AI Verification and Evaluation Research Institute, suggested that requiring frontier developers to contribute could ensure adequate funding for computing resources and expertise. “I think of these tests as public goods, but there are reasonable questions about if taxpayers should foot the bill,” Purser said.