NSA Spending Billions to Test Frontier AI Models for Security Vulnerabilities

The National Security Agency (NSA) is spending billions of dollars in taxpayer funds this year to evaluate and test advanced artificial intelligence models, according to classified intelligence estimates reported by The Washington Sun. This expenditure is significantly higher than previously known and has led lawmakers to estimate that a comprehensive federal AI regulatory system could cost the government tens of billions of dollars per year.

Drivers of High Costs

The primary cost drivers for the NSA’s Artificial Intelligence Security Center include:

• Computing Power: The largest expense is the processing power required to run and test frontier models. Global demand for chips has surged, making computing resources incredibly expensive. For context, Anthropic has secured computing deals potentially worth $517 billion, according to The Information.
• Personnel: Hiring top AI engineers is costly, with private sector salaries reaching hundreds of millions of dollars. The NSA struggles to compete with these packages despite having deep technical expertise.

Regulatory Implications

This spending highlights the financial burden of independent AI safety reviews. Previous Congressional Budget Office (CBO) estimates for specific regulatory proposals were much lower:

• The AI Security and Innovation Act was estimated to cost roughly $20 million per year.
• A separate bill for AI reporting and tracking was estimated at $36 million over five years.

The high actual costs have intensified debates over who should pay for AI safety. While some tech leaders like Elon Musk suggest frontier labs should self-regulate, experts argue this lacks independence. Alternatives include levying a tax on AI companies to fund independent audits, a model Anthropic and OpenAI have reportedly indicated they might support. Currently, the NSA’s efforts are funded by classified portions of the national security budget.