NSA Reportedly Spends Billions Testing Frontier AI Models

Summary

A report based on two people familiar with classified intelligence estimates says the NSA is spending billions of dollars this year testing and evaluating frontier AI models for national-security vulnerabilities. No public budget document has confirmed the figure, the Pentagon has not commented, and the claim rests entirely on anonymous sourcing. The spending is reportedly connected to the NSA’s Artificial Intelligence Security Center, but that connection, the precise scope of the work, and the evaluation program itself have not been confirmed by public NSA, Defense Department, or congressional records. The reported costs are attributed mainly to the computing infrastructure needed for repeated adversarial tests and to specialist personnel. Private-sector AI compensation can reach tens or hundreds of millions of dollars, making recruitment difficult for government agencies even though the NSA has a substantial technical workforce. The report contrasts the alleged NSA spending with the $20 million authorized for each fiscal year from 2027 through 2032 under the proposed civilian NIST center in H.R. 9363, alongside an estimated $80 million in implementation costs from 2026 through 2031. The two efforts would have different purposes: classified national-security testing versus civilian AI risk measurement and standards. The disparity has reportedly prompted lawmakers to consider charging frontier AI developers fees to fund independent evaluations. Anthropic and OpenAI have reportedly expressed openness to federal oversight, but the available reporting does not show that either agreed to finance it. Anthropic, Google, and OpenAI have also discussed a shared AI safety standards body, with OpenAI confirming the talks and saying industry standards should complement, not replace, federal safeguards and democratic oversight. The proposed body’s charter, governance, and funding remain unidentified, leaving unresolved whether independent oversight will be government-built, developer-funded, or under-resourced.