The Pentagon officially placed it on a list of security risks. The White House ordered federal agencies to stop working with it. And yet today, this very company is helping U.S. intelligence officers find weaknesses in foreign computer systems.
Anthropic, one of the most prominent American companies developing artificial intelligence, has found itself in quite a predicament. On one side, a legal dispute with the military. On the other, a secret contract with intelligence agencies. And in the middle of it all is a model code-named Mythos, which is so capable that the company refused to release it on the open market.
The Defense Department's Dispute with Anthropic
The entire conflict began with a seemingly simple request. The Department of Defense wanted assurances from Anthropic that it would be able to use its technology for "any lawful purpose." The company refused and clearly stated that its models must not power autonomous weapons systems or be used for the mass surveillance of American citizens.
That was unacceptable to the Pentagon. Defense Secretary Pete Hegseth labeled Anthropic a "supply chain risk," a designation that had never before been applied to an American company in U.S. history. The White House then issued an order instructing federal agencies to stop using Anthropic's technology. Trump wrote on social media that they "will not do business" with the company.
Anthropic responded with a lawsuit. A federal judge ruled in the company's favor and described the Pentagon's decision as "Orwellian" and unconstitutional.
The Problem No One Expected
While lawyers battled in court, another problem was growing behind the scenes. Intelligence agencies, including the NSA and CIA, were confronting a harsh reality: they do not have enough computing power to run the latest artificial intelligence models on their top-secret networks. A chip shortage that experts had not anticipated just a few years ago had grown into a security problem. The intelligence community was testing and deploying AI tools more slowly than it needed to. And China was closing in from behind.
The White House approved a classified request for nine billion dollars to purchase cutting-edge chips for intelligence agencies. Congress must formally approve the package, but in the meantime, approximately one billion dollars was redirected from other government budgets to immediately ramp up purchasing capacity.
The White House Circumvents Its Own Ban
White House Chief of Staff Susie Wiles did what makes sense within the logic of Washington politics. She quietly authorized the NSA to continue using Anthropic's advanced model, despite the Pentagon still formally considering the company a threat. The reason was simple. Anthropic is currently the most capable option available to intelligence officers. And in its competition with China, the U.S. cannot afford to wait.
At the same time, a classified contract was being negotiated. It contains the key condition that originally sparked the entire dispute: AI models must not be deployed on data belonging to American citizens. The Pentagon had rejected this clause and demanded a free hand for "any lawful purpose." The new contract does not include that provision. The White House also indicated that it wants to use this agreement as a model for cooperation with other AI companies.
Mythos: A Model Too Powerful to Be Public
The central element of the entire story is a model called Mythos. Anthropic introduced it in April 2026 as a tool specifically designed for cybersecurity, but also announced that it would not make it commercially available. The reason? The model is so capable at finding and exploiting security weaknesses that it could cause enormous damage in the wrong hands.
Mythos has reportedly found thousands of high-severity vulnerabilities, including flaws in every major operating system and every web browser. Only about fifty selected partners were given access to it, roughly a dozen of which have been publicly confirmed. The UK's AI Security Institute confirmed that it is among them. The NSA apparently is as well. Intelligence officers use it primarily to scan environments and identify exploitable vulnerabilities. In other words, exactly what it was designed to do.
The Dispute Persists
In April, Anthropic CEO Dario Amodei met at the White House with Susie Wiles and Treasury Secretary Scott Bessent. The meeting was described as productive. When asked about the meeting, Trump replied that he had "no idea" about it.
AI analyst Dean Ball, who was behind the first version of the Trump administration's AI agenda, summed it up: Trump backs down from people who can hit back, and Anthropic is that kind of company.
Anthropic's case illustrates a tension that will continue to grow. How far can a private company go in setting the conditions under which the government may use its technology? And what happens when that company is indispensable at the same time?
Edward Snowden, the NSA whistleblower, warned about this scenario years in advance. It is not enough for a company to publicly say "no" to mass surveillance. That refusal must be built directly into the architecture. Otherwise, everything depends solely on the goodwill of the people sitting at the table. Amodei gave his models something like a "resignation button." They can refuse to carry out an instruction if it conflicts with their configured values. The NSA will test that at every turn.
Sources: yahoo.com, nytimes.com and theinformation.com
