NSA Loses Access to Anthropic’s Mythos AI Model Over Export Dispute

In June 2026 the National Security Agency (NSA) lost access to Anthropic’s Mythos AI model after the Trump administration imposed export controls on the startup’s models, Mythos 5 and Fable 5. The controls forced Anthropic to suspend the models, ending a Project Glasswing partnership that let NSA analysts probe classified networks for software flaws. Anthropic’s refusal to let the U.S. military use its models prompted government’s blacklist.

Reuters

• June 11, 2026 – Senator Warner highlighted Mythos at a Senate Banking Committee hearing.

Reuters

• June 23, 2026 – Reuters reported Mythos found vulnerabilities in classified U.S. systems during Project Glasswing testing; the same day the Trump administration imposed export controls on Anthropic’s models and blacklisted the firm.

Reuters

• Late June 2026 – NSA analysts lost access to Mythos after Anthropic withdrew the models.

Reuters

Project Glasswing tests showed Mythos could locate flaws within hours, a pace the team called “a speed that human teams would struggle to match.” The model was reported to have breached nearly all classified systems in a matter of hours.

The AI model that broke into NSA systems is now the one the NSA cannot use

The NSA’s loss of Mythos removes a rapid-assessment tool that could speed remediation of vulnerabilities across classified networks, widening the security gap with adversaries. The episode also shows how export-control policy can limit frontier AI use for national-security work, raising oversight versus capability questions.

The AI model that broke into NSA systems is now the one the NSA cannot use

Sen. Mark Warner told the Senate that NSA chief Gen. Joshua Rudd briefed him on Mythos’s breach of classified systems. White House and Anthropic declined comment; the Defense Department referred queries to the NSA. The administration defended the export controls as necessary for national security.

Reuters

Anthropic’s refusal to let the U.S. military use its models led to its placement on a blacklist. An AP official stressed that identifying vulnerabilities does not equal exploiting them, warning against overstating the model’s offensive potential.

Reuters

Sen. Warner’s briefing said Mythos “broken into” classified systems, while the AP clarified the model only identified weaknesses without exploiting them. No findings have been released, and White House and Anthropic have offered no further explanation for the export-control decision.

The AI model that broke into NSA systems is now the one the NSA cannot use

• “broke into almost all of our classified systems, not in weeks, but in hours.” — Sen. Mark Warner, citing Gen. Joshua Rudd

New York Times

• “official, the AP said that although Mythos identified certain vulnerabilities within hours, that did not mean the model was able to exploit them within that time.” — Unidentified US official, AP

Reuters

Congressional committees will hold hearings on AI export controls, and the administration expects to issue guidance on frontier-AI use for intelligence. Anthropic has signaled willingness to resume cooperation under revised terms, while the NSA evaluates alternative rapid-assessment tools.

The AI model that broke into NSA systems is now the one the NSA cannot use