Anthropic’s Mythos found flaws in classified US systems during a government test

Anthropic's Mythos AI model identified vulnerabilities in classified US government systems during a controlled red-team exercise, not an external breach. The model surfaced flaws within hours, consistent with earlier evaluations where it found thousands of zero-day vulnerabilities across major OSes and browsers. The finding came to public attention via a Senate hearing where Senator Mark Warner quoted NSA/Cyber Command leadership. The episode highlights a contradiction: the same US government that uses Mythos on classified networks has also restricted it publicly following a separate jailbreak dispute. The UK's AI Security Institute independently assessed Mythos as substantially more capable at cyber offence than any previously tested model. Senator Warner cited the test to argue for mandatory pre-release evaluation of frontier AI models, not to condemn Anthropic.