Anthropic built an AI called Mythos. Mythos hacked classified government systems. Not attempted to hack. Hacked. Found vulnerabilities in the computers that house secrets we allegedly kill people to protect.
The official who announced this sounded proud. That's the detail worth examining. Someone stood in front of reporters and said yes, our AI found holes in the systems that defend nuclear codes and surveillance networks and whatever else lives behind "highly sensitive" clearance levels. Then everyone nodded like this was progress.
Two possibilities exist here. First: the vulnerabilities were already there and Mythos just found them faster than a human pentester would. Which means classified systems had exploitable flaws and we needed a chatbot to point them out. Second: Mythos created novel attack vectors no human had conceived. Which means we taught a language model to invent ways to breach national security infrastructure then told the press about it.
Either scenario suggests someone f*cked up in a way that would normally end careers. But this is AI coverage in 2026 so the frame is innovation.
Anthropic charges $30 per million tokens for Mythos. The Department of Defense has now confirmed their product can break into classified systems. Every foreign intelligence service with a credit card just received extremely specific product validation. China's MSS probably sent a thank-you card.
Retail traders will somehow find a way to lose money on this news. They'll pile into defense tech ETFs or AI cybersecurity penny stocks and get destroyed by theta decay on options they don't understand. Mythos will breach the Pentagon again next month and those same traders will act surprised their calls expired worthless.
The vulnerabilities have reportedly been patched. Until Mythos 2 finds the next ones.
Photo by on Unsplash
Leave a Comment