In February, President Trump ordered every federal agency to stop using Anthropic's AI services. Two months later, the National Security Agency is using Anthropic's most powerful model anyway.

According to Axios, which spoke to two sources with direct knowledge, the NSA is one of roughly 40 organizations Anthropic granted access to Mythos Preview, the AI model the company describes as "strikingly capable at computer security tasks." One source said the model is "being used more widely within the department" beyond just the NSA.

Let that sink in. The same administration that labeled Anthropic a "supply chain risk" and tried to blacklist it from government contracts is now running its most sensitive cybersecurity operations on Anthropic's most dangerous model. The left hand does not know what the right hand is doing, and the right hand is too busy stopping cyberattacks to care.

The Timeline That Makes This Absurd

Here is the sequence of events: In February, Trump banned all federal agencies from using Anthropic after the company refused to remove safety guardrails for military applications. In March, the Pentagon designated Anthropic a supply chain risk. Anthropic sued in two courts. One judge granted a preliminary injunction blocking the designation. Another denied it.

In early April, Anthropic launched Mythos Preview, a model so capable at finding and exploiting software vulnerabilities that five central banks held emergency meetings about it. The Treasury Secretary and Fed Chair called an emergency session. Bloomberg asked whether a single AI model could crash the global banking system.

And while all that was happening, the NSA quietly signed up to use it.

The Government Cannot Quit Anthropic

This is not the first crack in the ban. We have been tracking federal agencies quietly testing Anthropic models for weeks. But the NSA is different. This is the crown jewel of American signals intelligence, the agency responsible for protecting the country's most classified communications, openly defying a presidential directive because the cybersecurity need is that urgent.

Last week, Anthropic CEO Dario Amodei walked into the White House for a meeting with Chief of Staff Susie Wiles. The White House called it "productive and constructive." When reporters asked Trump about the meeting, he said he had "no idea" it happened. Sources inside Anthropic have said privately that refusing to give the US government access to Mythos would be "grossly irresponsible" and "a gift to China."

The argument is winning. Not in courtrooms. Not in policy meetings. On the ground, where the people responsible for defending American infrastructure are making pragmatic choices regardless of what the White House says publicly.

What This Actually Means

The ban is dead in practice, even if it is alive on paper. When your own intelligence community decides it needs the tool more than it needs to follow your policy, the policy has failed. The question now is not whether the administration will reverse the ban. It is when, and how they will frame the reversal to avoid looking like they lost.

Yoshua Bengio, the AI pioneer, criticized Anthropic this week for keeping Mythos locked to only 40 organizations, arguing it concentrates decision-making power within a single company. Yann LeCun called the whole cybersecurity narrative overblown. But neither of those arguments matter to the NSA director who just watched Mythos find zero-day vulnerabilities that human teams missed.

The Anthropic saga has been the defining story of 2026 AI policy. The company that said no to the Pentagon is winning by being too useful to ban. The irony is thick enough to cut: the model Trump tried to kill is now protecting his own government's networks.

First reported by Axios. Confirmed by Engadget with additional context.