Export Controls Cut NSA Off From Anthropic AI Models Tested on Classified Systems

The National Security Agency lost access to Anthropic's two most advanced artificial intelligence models after the Commerce Department issued an export control directive on June 12, 2026, ordering the company to suspend access to Fable 5 and Mythos 5 for any foreign national, including Anthropic's own non-citizen employees ^[1][2]. Unable to enforce a nationality-based access screen on its user base, Anthropic disabled both models globally, cutting off government agencies and private customers simultaneously ^[3][4]. The action marks the first time the United States has applied export controls directly to an AI model rather than to the hardware or chips powering it, establishing a regulatory precedent with broad implications for the AI industry ^[5].

The directive arrived as the NSA was actively evaluating Mythos 5 for cybersecurity applications. Parts of the NSA lost access to Mythos 5 following the export control action, with some analysts notified on a Friday that they would lose access. The agency may retain limited access to earlier versions of the technology under prior arrangements, though access to company support, updates, or modifications is now more constrained. The NSA had been able to continue testing Anthropic's models even after the Pentagon designated the company a supply-chain risk in March, according to The New York Times [POLITICO]. NSA's access flowed through Project Glasswing, Anthropic's restricted early-access program for security researchers and government-linked organizations, which Anthropic had expanded to roughly 150 organizations in more than 15 countries.

The export controls followed public statements by Sen. Mark Warner, D-Va., vice chairman of the Senate Intelligence Committee, who disclosed that NSA Director Gen. Joshua Rudd had told him Anthropic's Mythos model "broke into almost all of our classified systems, not in weeks, but in hours." That characterization spread rapidly online before officials clarified its scope. The Economist's defense editor later posted on social media that a U.S. official told him Warner had misunderstood Rudd's comments and that the Mythos work was part of a red-teaming effort to test the security of internal systems. The exercise was an authorized internal red-team test in which Mythos was paired with other defensive tools under highly specific simulated environmental conditions. Officials nonetheless told the Times that NSA analysts were struck by Mythos's performance in finding software vulnerabilities within that controlled setting ^[4].

The export control directive itself was triggered in part by a reported jailbreak of Fable 5. The government said a trusted partner, which CNN identified as Amazon, found a jailbreak, or a way to circumvent Fable's safety guardrails. Anthropic said it disagreed with the government's handling of the matter, stating it received the directive Friday afternoon and that it did not specify the national security concerns. Anthropic contended the White House decision was based on "one potential jailbreak" shared by Amazon CEO Andy Jassy, according to The Wall Street Journal. The Commerce Department's Bureau of Industry and Security issued the directive under export control authority; the White House referred questions to President Trump's recent cybersecurity executive order, which invites AI companies to submit frontier models for government security review on a voluntary basis [POLITICO]^[7].

The export controls land against a backdrop of protracted litigation and regulatory conflict between Anthropic and the federal government. The Pentagon officially designated Anthropic a supply-chain risk in early March, and a federal judge in San Francisco later granted Anthropic a preliminary injunction in a separate case barring the Trump administration from enforcing the broader ban on Claude's use. A federal appeals court in Washington subsequently denied Anthropic's request to temporarily block the Pentagon's blacklisting while that lawsuit proceeds. Anthropic is the first American company to receive the supply-chain risk designation, which has historically been reserved for foreign adversaries. The root dispute involves Anthropic's refusal to remove contractual restrictions barring its Claude models from use in mass domestic surveillance or fully autonomous weapons systems. In July 2025, Anthropic and the Pentagon entered a contract making Claude the first frontier model approved for use on classified networks, with the Pentagon agreeing to abide by Anthropic's acceptable use policy, which prohibited use for mass surveillance of Americans or in fully autonomous weapons systems capable of selecting and engaging targets without human intervention.

The timing of the export control dispute sharpens a tension already flagged by the intelligence community. On the same week, the Five Eyes intelligence alliance warned that frontier AI models could sharply change the cyber threat landscape within months, not years, by helping attackers and defenders move faster. A classified contract that would permit the NSA to use Anthropic technology for intelligence analysis and vulnerability detection has not been finalized, and some Pentagon officials favor transitioning to other models. As of press time, negotiations between Anthropic and the government remained unresolved. The Commerce Department lifted the export controls on June 30 after Anthropic implemented new safeguards to address the reported jailbreak vulnerability, with the company announcing it would begin restoring access to Fable 5 and Mythos 5 the following day ^[6][7].