Anthropic's latest AI model, called Mythos, reportedly breached "almost all" of the NSA's classified systems during a controlled red-team test, finishing the job within a few hours.
According to Tom's Hardware, the claim comes from a quote cited by The Economist. A red-team test is a security exercise in which a system is deliberately attacked under controlled conditions to find weaknesses before real adversaries do — in this case, the AI was apparently turned loose against government networks to see how far it could get.
Tom's Hardware frames the report as adding context to a previously unexplained move: the U.S. government's sudden decision to restrict, or ban, Anthropic's flagship models. The reporting suggests the red-team result may help explain why officials acted so abruptly.
It is worth being precise about what is known here. The detail is described as reportedly true and rests on a single quote that The Economist cited and that Tom's Hardware relayed. The exact scope of "almost all," the conditions of the test, and Anthropic's own account are not detailed in the source.
Why it matters: if a commercially available AI can penetrate hardened classified systems this quickly, it signals that the same capabilities companies are selling to businesses could reshape both cyberattack and cyberdefense — and helps explain why governments are moving fast to control which models they allow near sensitive networks.