An Axios scoop published April 19 says the National Security Agency is already using Anthropic's most capable model, Claude Mythos Preview, even as the Department of Defense — the NSA's parent agency — continues to brand Anthropic a 'supply chain risk.' The report, citing unnamed sources, claims Mythos Preview is 'being used more widely within the department.' Reuters said it could not independently verify the account, and Anthropic, the NSA, and DoD did not immediately respond to requests for comment.
The story drops into the middle of one of the most extraordinary standoffs in the short history of frontier AI procurement: a vendor the Pentagon has formally tried to exclude is, according to the report, actively servicing one of the most powerful agencies the Pentagon oversees.
How Anthropic ended up on the blacklist
The rift traces back to contract renegotiations earlier this year. The DoD reportedly demanded Anthropic make Claude available for 'all lawful purposes,' a category broad enough to encompass mass domestic surveillance and autonomous-weapons development. Anthropic refused to drop those carve-outs and, in February, the department formally designated the company a supply chain risk and pushed its vendors to follow suit.
That would normally close the door across the federal government. It did not.
Why the NSA wants Mythos anyway
Mythos Preview is the model Anthropic has so explicitly classified as too dangerous for general release that it built an entire access regime around it. Under Project Glasswing, around 40 additional organizations — on top of 12 launch partners including AWS, Apple, Google and Microsoft — get gated access, primarily to use the model's offensive security capabilities for defense. Anthropic has described Mythos as 'a new general-purpose language model' that is 'strikingly capable at computer security tasks,' and security researchers have warned the same skills that find vulnerabilities can be used to weaponize them.
For an agency whose remit includes signals intelligence and cyber defense, that capability profile is the entire point. The reported deployment suggests the NSA's operational need outweighed the political optics of running software from a blacklisted vendor.
A White House thaw
The Axios report lands two days after Anthropic CEO Dario Amodei met with White House Chief of Staff Susie Wiles and Treasury Secretary Scott Bessent on April 17. Both sides reportedly described the meeting as productive, and prior reporting indicated the White House has been pushing to make Mythos accessible across federal agencies despite the Pentagon dispute.
What it signals
The episode hardens a pattern that has been visible for months: the Pentagon's procurement levers are no longer the single chokepoint for frontier AI inside the U.S. government. When a model is good enough — and scarce enough — agencies are finding paths to it that route around the blacklist. Whether the DoD now escalates, quietly relents, or watches its own designation become a dead letter is the next thing to watch.
