Axios exclusive: The U.S. NSA bypassed the Pentagon blacklist to use Anthropic Mythos, and Dario Amodei urgently met with the White House to negotiate

According to an Axios exclusive report and relayed by Reuters, the U.S. National Security Agency (NSA) is using Anthropic’s strongest model, Mythos Preview, even though its parent agency—the Department of Defense (DoD)—since February has classified Anthropic as a “supply-chain risk” and banned it. This is the first public contradiction in U.S. AI governance history where one department of the same federal government blocks it while another department is actually using it.

Mythos is open to 40 official organizations; only 12 are publicly listed

After Anthropic publicly revealed the existence of Claude Mythos in February, it granted preview access to only about 40 organizations worldwide, but the official publicly released list included just 12. The rest were deliberately concealed based on the rationale that “offensive cyber capabilities are too sensitive.” As the core unit within the U.S. intelligence community that handles signal intelligence (SIGINT) and cryptography, the NSA is not among Anthropic’s publicly listed 12, yet it has actually obtained and is using Mythos—prompting double scrutiny of both Anthropic’s list transparency and the U.S. government’s access authorization.

A self-contradiction: the Pentagon’s ban vs. the military’s expanded use

The relationship between the Pentagon and Anthropic broke down earlier this year. The official position was that, during contract updates, Anthropic insisted on “excluding uses such as large-scale domestic surveillance and autonomous weapons development,” while the Department of Defense required Mythos to be “open for all lawful purposes.” After negotiations collapsed, in February the DoD added Anthropic to the disabled list on the grounds of “supply-chain risk,” and required defense suppliers to cut ties as well.

Ironically, two months after DoD formally disabled it, not only did the NSA continue using Mythos, but multiple branches of the U.S. military even expanded adoption of other Anthropic Claude series models. Government entities have publicly argued in court filings that “using Anthropic tools would threaten national security,” yet the same entities continue internally to use Claude to handle cybersecurity and intelligence missions. Earlier this month, Treasury Secretary Bessent and Federal Reserve Chair Powell held an emergency meeting with senior leaders at a bank to discuss regulatory concerns over Mythos. Now the internal contradictions have come to the surface, and the administrative risk for Mythos has risen significantly.

Dario Amodei goes to the White House to negotiate a truce

Anthropic CEO Dario Amodei went to the White House on April 17 to meet with Chief of Staff Susie Wiles and Treasury Secretary Scott Bessent. According to Axios, the talks focused on three issues: the boundaries of Mythos’s use within government systems; Anthropic’s overall security operations and review processes; and how to allow “departments and agencies other than the Pentagon” to continue obtaining access. The next steps afterward may include establishing separate authorization pathways for non-DoD federal agencies, or requiring Anthropic to disclose a more complete list of customers to respond to Pentagon regulatory concerns.

A new template for tiered access and government procurement boundaries

The Mythos incident, occurring in the same week as OpenAI’s launch of GPT-5.4-Cyber, echoes the same trend: in 2026, the frontier AI “strongest” model will enter a tiered release phase where the stronger the capability, the fewer customers. Such control frameworks appear strict, but when the U.S. federal government’s internal rules are inconsistent, in practice it amounts to the vendor (Anthropic) deciding which government units are trustworthy. For governments and companies in other countries, the next policy pressure is whether to require AI vendors to publicly disclose customer authorization lists when processing data involving their country’s citizens. The actions Anthropic takes over the next few weeks will become the governance template for the entire frontier AI industry.

This Axios exclusive: U.S. NSA bypasses the Pentagon blacklist to use Anthropic Mythos; Dario Amodei urgently goes to the White House to negotiate
First appeared on Chain News ABMedia.