NSA Is Using Anthropic's Restricted Mythos Model Despite the Company's Feud With the Pentagon

The NSA is quietly working with Mythos, a restricted Anthropic AI model, despite the company's public standoff with the Pentagon. According to TechCrunch, the US National Security Agency has already integrated the restricted model into its intelligence work, even though relations between the developer and the military are officially strained. The outlet's sources say this is not a pilot test but real-world use.

Mythos is a restricted Anthropic model not available to the general public. Unlike Claude, which is distributed through an API, subscriptions and enterprise contracts, Mythos lives in a separate, closed environment. Details about the model's capabilities, training methods, system prompts and licensing terms have not been disclosed publicly. Until recently, Anthropic had not officially confirmed that Mythos exists.

Analysts suggest several possible ways Mythos differs from standard Claude. Possible explanations include extended context windows for long analytical documents, extra accuracy on the specific formats of intelligence reports, and relaxed restrictions on certain types of queries relevant to the public sector. An alternative reading is that "restricted" simply means controlled distribution: the model is not sold openly but supplied only to vetted government partners.

Anthropic has consistently positioned itself as a "responsible" AI developer: the company promotes safe-development principles and professes a cautious approach to military applications. That has created a noticeable public distance between Anthropic and the defense sector, unlike Palantir, Anduril or Scale AI, which openly build their businesses on Department of Defense contracts.

Still, the line between "civilian" and "government" AI providers is thinner than commonly assumed. The intelligence community operates under a different chain of command than the Pentagon: the NSA reports to the Director of National Intelligence, not the Department of Defense. Formally, that lets the agency sign contracts with vendors regardless of the military's position, even if a company publicly distances itself from "military" uses.

AI in US intelligence work is hardly news. The CIA launched an internal tool built on large language models back in 2023 and by 2025 had expanded cooperation with several AI companies. The NSA has traditionally been ahead of the public discourse in adopting new technology. For intelligence agencies, the key AI uses are processing huge volumes of intercepted signals, automating pattern detection in communications and speeding up analytical reports. In that context, a powerful language model in the agency's arsenal is a logical step rather than a sensation.

The situation puts Anthropic in front of a contradiction familiar across the tech sector: public ethical stance and actual commercial practice do not always match. For a company actively building an image as a "safe" developer in contrast to more aggressive market players, reports of cooperation with the country's main spy agency are a reputational challenge.

Neither Anthropic nor the NSA has officially commented on the outlet's report. While the public narrative of the company's feud with the Pentagon lives its own life, real business relations with the intelligence community appear to be carrying on as usual.