The U.S. National Security Agency (NSA) is reportedly using Mythos Preview, a recently introduced AI model from Anthropic that was not released to the public, according to Axios. The development comes just weeks after the U.S. Department of Defence's military oversight body labelled Anthropic's parent a "supply-chain risk," escalating tensions between the two sides.
Anthropic unveiled Mythos earlier this month as a frontier-level model built specifically for cybersecurity applications. However, the company said the system was too capable of generating offensive cyberattack techniques to be safely released publicly. As a result, access to Mythos has been restricted to roughly 40 organisations, with only a small number publicly disclosed.
While Anthropic has not officially confirmed all users, reports suggest that the NSA is among the select group granted access. The agency is said to use Mythos primarily to identify and analyse potential security vulnerabilities across systems. The U.K.'s AI Security Institute has also confirmed it has access to the model.
The situation highlights a growing contradiction in the U.S. government's approach to advanced AI tools. On one hand, military and intelligence agencies are increasingly adopting Anthropic's systems for cybersecurity and defensive operations. On the other hand, parts of the Pentagon have raised concerns about the same technology, arguing in legal disputes that unrestricted access could pose national security risks.
The original conflict between Anthropic and the Pentagon reportedly stemmed from negotiations over how the military could use the company's models. Anthropic resisted requests to enable use cases involving mass surveillance and fully autonomous weapons systems, prompting the Pentagon to classify the company as a potential supply chain risk.
Meanwhile, Anthropic's relationship with the broader U.S. administration appears to be stabilising. Last Friday, Anthropic CEO Dario Amodei met with White House Chief of Staff Susie Wiles and Treasury Secretary Scott Bessent, in discussions officials described as constructive.
The evolving situation reflects a broader debate inside government and industry circles over how powerful AI systems should be controlled, who should have access to them, and how they should be used in national security contexts.
