Anthropic AI model triggers US export ban after rapid NSA security test

A security test of Anthropic's advanced AI models, including Mythos and Fable, identified vulnerabilities in U.S. classified computer systems within a few hours. The test, conducted in cooperation with the National Security Agency (NSA) and the U.S. Cyber Command, was described by a senior official as showing the tool "penetrated almost all of our classified systems, not weeks but hours."

Following the test, the Trump administration issued an export‑control order requiring Anthropic to suspend foreign citizens' access to its latest models, citing national‑security concerns. Anthropic said the exercise was an authorized internal simulation and that no actual breach occurred, and it protested the lack of a formal justification for the restrictions. The company also announced it would block foreign use of its models worldwide.

Senator Mark Warner publicly referenced the test, and industry leaders from firms such as Adobe and Nvidia signed letters urging the government to reverse the order, warning it could advantage U.S. rivals. The dispute highlights growing tensions between AI developers and policymakers over how to manage potential security risks of powerful artificial‑intelligence systems.