NSA Is Using Anthropic's Powerful Claude Mythos AI as CEO Meets With White House: Report
The NSA is reportedly running Anthropic's Claude Mythos Preview on classified networks, even as the Pentagon fights the AI giant in court.
By Jose Antonio LanzEdited by Andrew HaywardApr 20, 2026Apr 20, 20263 min read
In brief
- The NSA is deploying Claude Mythos Preview despite the Pentagon—which oversees the agency—designating Anthropic a supply-chain risk in March, per a report.
- Anthropic CEO Dario Amodei met with White House Chief of Staff Susie Wiles and Treasury Secretary Scott Bessent on April 17, with both sides calling the discussions "productive."
- An administration source told Axios that every federal agency except the Department of Defense wants access to Anthropic's AI tools.
The National Security Agency is running Anthropic's Claude Mythos Preview inside its classified networks, according to two sources cited by Axios—a surprising development given that the NSA falls under the Department of Defense, which declared Anthropic a supply-chain risk in March and is currently fighting the company in federal court.
Claude Mythos is not a standard enterprise tool. When Anthropic unveiled the model earlier this month, it restricted access to a handful of vetted organizations, arguing that the model poses serious offensive security risks. Anthropic's own technical documentation found that Mythos was able to identify critical vulnerabilities in every widely used operating system and web browser. The company judged it too dangerous for open release.
Most organizations with access are using the model defensively, scanning their own infrastructure for weaknesses before adversaries do. The initiative, branded Project Glasswing, includes Microsoft, Google, Apple, Amazon Web Services, JPMorgan Chase, and Nvidia. What the NSA is doing with Mythos is less clear, though the agency's mission is not purely defensive. A third source told Axios the model is being used more broadly within the intelligence department.
The Pentagon's hostility toward Anthropic traces to negotiations that went bad. In July 2025, the two sides signed an agreement making Claude the first frontier AI model cleared for use on classified networks. Talks soured when the Pentagon sought to renegotiate, demanding the military be allowed to use Claude "for all lawful purposes" without restriction. Anthropic refused, drawing two firm lines: no autonomous weapons, and no domestic mass surveillance.
When negotiations collapsed, Defense Secretary Pete Hegseth declared Anthropic a supply-chain risk in late February—an unprecedented designation, and the first ever applied to an American company. A California federal judge blocked the move, but then a D.C. appeals court denied Anthropic's separate bid to halt the blacklisting while litigation plays out. The two sides remain in court.
While the legal fight grinds on, the rest of the administration is moving in a different direction. On April 17, Anthropic CEO Dario Amodei met with White House Chief of Staff Susie Wiles and Treasury Secretary Scott Bessent. Anthropic described the session as “productive”, Reuters reported. The White House said the parties "discussed opportunities for collaboration, as well as shared approaches and protocols to address the challenges associated with scaling this technology."
President Trump, asked by reporters about the meeting, said he had "no idea" Amodei had been at the White House, after he previously ordered the administration not to use Anthropic’s “woke” models. Bessent and Federal Reserve Chair Jerome Powell have separately been encouraging major bank CEOs to test Mythos and be prepared for security threats, and an administration source told Axios that every federal agency except the Defense Department wants access to Anthropic's tools.
The NSA's reported use of Mythos comes as questions mount about whether the model's capabilities can be contained at all. Decrypt reported last week that researchers at Vidoc Security reproduced several of Mythos's most alarming cybersecurity findings using publicly available models—including OpenAI’s GPT-5.4 and Anthropic’s own Claude Opus 4.6—without any special access to Mythos itself.
Anthropic did not immediately respond to a request for comment by Decrypt.
