NSA Spending Billions to Test AI Models, Classified Estimates Show
Classified budget figures indicate the NSA has become one of Washington's biggest payers for AI evaluation, funding red-teaming and security testing at a scale rarely seen outside the intelligence community.
The National Security Agency is spending billions of dollars to evaluate artificial intelligence models, according to classified budget estimates reported by the Washington Sun. The figures, which have not been confirmed by the agency, put a dollar amount on a slice of the intelligence budget that has stayed almost entirely out of public view. If they hold, model testing has quietly grown into one of the NSA's larger technology line items.
The money covers assessment, not acquisition. That distinction matters: the reported spending goes toward red-teaming, benchmarking, and probing systems for failure modes, bias, and security holes before they are cleared for classified networks. Running that kind of work at the scale the estimates imply points to a standing review pipeline rather than one-off vetting of a handful of tools.
The NSA has said little about how it adopts commercial AI. Its known efforts run through the AI Security Center, established in 2023 under then-director Gen. Paul Nakasone to coordinate AI security across the defense and intelligence communities. The center's remit includes tracking adversary use of AI and hardening U.S. systems, work that depends on exactly the continuous testing the reported budget would fund.
The numbers arrive without the documentation that usually accompanies figures of that size. The intelligence community's fiscal 2024 budget request topped $100 billion across the National Intelligence Program and Military Intelligence Program, but AI line items are rarely declassified, and outside analysts have no independent way to verify the estimates. The NSA has not confirmed them, and the classified annexes that would settle the question remain closed.
For the AI industry, the signal is that the federal government's largest customer for model evaluation may be an agency that never announces its purchases. Contractors doing security testing, red-teaming, and model auditing now have a plausible federal market measured in the billions, even if the awards stay classified. That could pull evaluation talent and tooling toward defense work and away from purely commercial customers.
What to watch next is whether any of this becomes visible. Contract awards, hiring activity at the AI Security Center, and congressional pressure to force more disclosure on intelligence spending for AI would all offer clues about what the money actually buys. Until then, the size and shape of the program rest on estimates no one outside the cleared community can check.