No alternative in sight: NSA taps Anthropic's Mithos for offensive cyber operations

Even as Anthropic finds itself at odds with the Donald Trump administration, use of the company's AI model Mithos across the administration continues to grow.

The National Security Agency has decided to deploy Mithos in offensive cyber operations, the Financial Times reported. Anthropic is said to have sent six engineers to the NSA to customize the model's capabilities and transfer operational know-how. Sources familiar with the NSA's contract with Anthropic said Mithos could eventually be used to penetrate the networks of countries such as China and Iran.

A person who spoke to the FT said "the best way to build a good defense is to build a good offense," adding that adversary nations are likely developing their own AI-based offensive capabilities. "If it's not used to build offensive agents, adversaries will find ways" to attack US networks, the person said, adding that the NSA's offensive cyber preparations are a direct response to that threat.

The NSA's adoption of Mithos stands out because it comes amid an ongoing clash between the Trump administration and Anthropic. The company has pushed back against the administration's use of its Claude AI model for immigration enforcement surveillance and for autonomous drone systems capable of carrying out lethal strikes. The Defense Department responded by taking the unusual step of designating Anthropic — a domestic company — as a "supply chain risk," prompting Anthropic to file a lawsuit.

Demand for Mithos has nonetheless kept rising. Anthropic unveiled the model in April, after the dispute with the Pentagon erupted, and it quickly made waves in the industry by demonstrating the ability to detect and exploit software vulnerabilities. Governments, financial institutions and technology companies have reacted with a mix of concern and interest, and the model is now in use at 150 organizations across 15 countries.