The AI That Taught Itself to Hack — And What Happens When China Gets It

Anthropic built an AI so dangerous they refused to release it. Mythos taught itself to find vulnerabilities, chain them together, and take down entire systems — autonomously, in minutes. No human in the loop. No patch fast enough to stop it. By Anthropic's own assessment, left unchecked it could dismantle a Fortune 100 company and disrupt large parts of the internet before your IT team finishes their morning coffee. So instead of releasing it, they handed it to 12 companies — Apple, Google, Microsoft, JP Morgan, and yes, the NSA — under a classified program called Glasswing. The idea: give the defenders a head start before attackers get the same capability. Here's the problem. Dario Amodei, Anthropic's own CEO, says open-source Chinese models will hit Mythos-level capability in 6 to 12 months. This was never really a technology story. It's a geopolitical one — and we're not sure the people in charge are equipped to handle it. Mike and Neal break down what Mythos actually does, why your power grid is the most terrifying target on the map, and how AI-powered phishing attacks are already hitting small businesses with click-through rates that should scare the hell out of you. They also get into where the real startup opportunity lives inside all of this — because defense tech is about to have a moment, and the window won't stay open forever. Plus: California's battery grid quietly hits a landmark milestone, the data center bubble shows its first real cracks, and the goods, bads, and others — including Neal's travel hell, a birthday, and Mike's feelings about the Baltimore Orioles. This is the stuff that's actually shaping your world. Let's get into it. Got a take? Email us: [email protected] Follow Hardpoints wherever you get your podcasts.