Why Federal Agencies Are Secretly Testing Anthropic’s AI Model Even After the White House Tried to Block It

When you stop accepting the official positions at face value, a certain type of Washington story starts to make sense. The Anthropic one appears to be a clear example. Defense Secretary Pete Hegseth and President Donald Trump ordered federal agencies to cease collaboration with the San Francisco-based AI firm in late February. The stated reasons were a combination of operational and political: the White House was unhappy about what it called “woke AI” disagreements, and Anthropic had refused to allow the Pentagon to use its technology in autonomous lethal systems or mass surveillance of Americans. The company was officially classified as a supply-chain risk by Hegseth. That was the end of it on paper.

on paper. Something strange is taking place in practice. The Commerce Department’s Center for AI Standards and Innovation has been actively testing Anthropic’s most recent model, Claude Mythos, over the past week, according to Politico, particularly for its capacity to identify software vulnerabilities. The outline was verified by Reuters. At least three congressional committee staff members have either requested or held briefings on the capabilities of the model. According to reports, Anthropic has received direct communication from at least two major federal agencies expressing interest in utilizing Mythos for cyber defense. This was not supposed to be taking place.

In this industry, the hype and the content don’t always align, so it’s worth stopping to consider mythos itself. Built specifically for coding and what the company refers to as “agentic tasks”—that is, the capacity to act somewhat independently rather than waiting for a human prompt at every stage—Anthropic unveiled it on April 7 as its most capable model to date.

Mythos scores significantly higher than Claude Opus 4.6 on software programming and academic reasoning benchmarks, according to internal drafts that were leaked due to a configuration error in Anthropic’s content management system that exposed about 3,000 unpublished assets. But it’s another aspect that national security officials are interested in. During testing, the model discovered software defects that human security researchers had overlooked. Not obscure bugs. serious ones.

There are two sides to that capability. By definition, a tool capable of identifying vulnerabilities that the most skilled engineers miss is also capable of identifying them first if directed in the wrong direction. For this reason, Anthropic is allegedly limiting early access to the model to a select group of cybersecurity defense clients and withholding it from the general public. Additionally, it appears that the Commerce Department determined that the risk of not testing a product from a banned company outweighed the political awkwardness of doing so. Reading between the lines of the reporting gives the impression that agency personnel sincerely think China and Russia are already developing equivalents. They don’t consider waiting for a White House order to be a luxury.

Here, the cultural narrative is nearly as fascinating as the technical one. Despite the Pentagon’s contract cutoff, Anthropic is still in contact with the Trump administration, according to co-founder Jack Clark, who spoke at the Semafor World Economy Summit on Monday. The language was cautious, formal, and a little worn out. It’s the tone of a business that has found that the government is both its primary client and its primary regulator, and that government hasn’t quite decided which it wants to be in any given week. In the meantime, a different Politico article from April 18 suggested that a ceasefire might be developing. It’s difficult to tell if that is a trial balloon or something real.

The underlying shape of the moment is difficult to overlook. A presidential ban couldn’t stop federal agencies from covertly testing something a private company had developed. That’s either a mythological tale or an account of how Washington functions in real life when technology advances more quickly than policy. Most likely both.