Anthropic's Mythos AI breached nearly all NSA classified systems in hours, senator says

Senator Mark Warner, vice-chair of the Senate Intelligence Committee, revealed that the general overseeing both the NSA and Pentagon Cyber Command told him Anthropic's Mythos model penetrated almost all classified NSA systems during a controlled red-team exercise. The breach occurred in hours rather than weeks, a speed that alarmed officials and prompted the U.S. government to restrict Anthropic's Mythos and Fable models to U.S. citizens only on June 12, 2026. Because Anthropic could not reliably verify user citizenship, it withdrew access entirely, affecting even close allied nations. The disclosure reframes the earlier restriction order: it was not a response to a content-safety violation but to a demonstrated offensive cyber capability deemed too dangerous to leave unrestricted. Experts note the test was sanctioned and controlled, but the AI's autonomous, rapid, self-correcting performance is what distinguished it from conventional human red-team efforts.