An Anthropic Mythos test with U.S. intelligence agencies reportedly found vulnerabilities in highly sensitive government systems within hours. The episode sharpens the policy problem around frontier AI: the same models that can help defenders fix critical software can also compress the timeline for attackers.