The Forbidden Model: Why the NSA is Secretly Using Anthropic’s Mythos AI Despite a Federal Ban

While the Trump administration recently mandated a comprehensive prohibition of Anthropic’s services across federal agencies—even designating the AI startup as a “supply chain risk”—the United States intelligence apparatus appears unable to resist the allure of its vanguard technology. According to Axios, citing individuals familiar with the matter, the National Security Agency (NSA) is currently utilizing Mythos Preview, a nascent AI model unveiled by Anthropic in early April.

Anthropic officially introduced Mythos Preview as a general-purpose language model, yet specifically emphasized its formidable prowess in computer security tasks. This characteristic strikes at the quintessential core of the NSA’s operational mandate. Sources indicate that Anthropic has restricted access to Mythos Preview to a select cohort of approximately forty organizations for preliminary evaluation, with the NSA among the chosen few. Internal reports further suggest that the model’s application within the Department of Defense is undergoing a rapid and extensive expansion.

The relationship between Anthropic and the federal government remains fraught with tension. In February, following Anthropic’s refusal to compromise on its “AI safeguards” regarding military applications, the Trump administration ordered all government entities to cease utilizing their services. However, the stance of high-ranking officials now appears to be softening. Last Friday, Anthropic CEO Dario Amodei was observed entering the White House for a summit with Chief of Staff Susie Wiles and other senior dignitaries, presumably to deliberate on the implementation of the Mythos model. The White House subsequently corroborated the meeting, characterizing the dialogue as “constructive and productive,” though President Trump, when queried by journalists, professed to be “entirely unaware” of the engagement.

Despite the clandestine rapport between the White House, intelligence agencies, and Anthropic, a fierce legal conflict persists. In March, Anthropic initiated two lawsuits against the Department of Defense to challenge the “supply chain risk” designation. The judicial outcomes thus far remain inconclusive: while one court granted a preliminary injunction to stay the label, another federal judge denied Anthropic’s motion to have it rescinded.

Superficially, the Trump administration must project a posture of political fortitude. When a Silicon Valley startup dares to demand that the Pentagon adhere to its ethical safety protocols during contract negotiations, Washington retaliates with heavy-handed administrative measures to deter such insolence. Yet, on the pragmatic battlefield of intelligence, the NSA and the Department of Defense recognize that supremacy in cyber warfare and cryptographic defense belongs to whoever wields the most potent analytical AI. Evidently, the performance of Anthropic’s Mythos model in the cybersecurity domain is so peerless that American intelligence agencies are willing to bypass executive orders under the guise of “early testing” to integrate it into their internal architectures.