The Department of Defense has been trying to get rid of Anthropic since March. It designated the company a supply-chain risk, ordered a six-month phase-out of Anthropic products across DoD, and is currently in court defending that decision against a lawsuit Anthropic filed in response. The Pentagon’s stated position, publicly and in legal filings, is that Anthropic is a problem.

Also, the NSA is using Anthropic’s newest model.

The NSA is part of the Department of Defense. Axios reported on April 19 that the agency has authorized access to Mythos Preview, the restricted frontier model Anthropic shipped earlier this month. The use case is scanning NSA environments for exploitable vulnerabilities, which is exactly what Mythos is good at, which is the reason Anthropic didn’t release it publicly. There are about 40 organizations on the Mythos allowlist worldwide. Most are not named. One of them is the signals-intelligence agency of a department that is simultaneously arguing, in federal court, that Anthropic cannot be trusted.

In principle, both things can be true. A model can be valuable enough that a responsible government wants its own spies to have it, and troublesome enough that the same government wants it out of PowerPoint-summarization contracts. Those are different problems. It is still a somewhat awkward look.

The timing helps. On April 18, one day before the Axios story, Dario Amodei met at the White House with Chief of Staff Susie Wiles and Treasury Secretary Scott Bessent. The White House described the meeting as “productive.” The White House did not describe what it was productive about. One reasonable guess is that it was productive about the fact that the Pentagon had picked a fight with a frontier-AI lab, and the rest of the administration would like that fight to be less visible, or at least less in the way of the AI the NSA would like to be using.

The Pentagon’s procurement strategy, described here last week when Google agreed to broadly the same terms Anthropic wouldn’t, is that labs take the standard contract or they don’t do classified work for DoD. That works well for tooling where any frontier lab is a substitute for any other. It works less well for capabilities where one company has a material lead and the operational need is now. Cyber defense is one of those. Mythos is, currently, one of those leads.

So the NSA has Mythos. Anthropic has a lawsuit, a White House meeting, and a private non-proliferation regime it runs itself, deciding on a per-customer basis which governments and agencies get the frontier’s most sensitive cybersecurity tool. No Chinese lab is on the list. DeepSeek is raising from Chinese capital only. The U.S. government is fighting in two directions about a company whose product it cannot, operationally, do without.

The blacklist is still in force. It is simply not the whole story. The NSA disclosure is the first clean admission of that.

Sources: Axios (original reporting), TechCrunch, BusinessToday, Slashdot discussion

–
By the Control Plane Editorial Team