What you'll learn

Project Glasswing is the first time a major AI lab has publicly stated it built a model too dangerous to release. That decision, and the closed coalition of around 40 organisations Anthropic has handed it to instead, tells you something important about where AI-augmented cybersecurity has arrived and how little time most organisations have to prepare.

This episode of Security Hot Takes works through the full picture. We examine what Claude Mythos Preview actually demonstrated in testing: autonomous vulnerability chaining, zero-days in hardened operating systems undetected for decades, and exploit conversion rates matching the most skilled human researchers. We look at the governance question this creates, who selected those 40 organisations and on what criteria, and what it means for the vast majority of the market that isn't in the room. We cover the dual-use reality at the heart of the initiative and what it demands of defenders right now. And we examine why the six-to-eighteen month timeline for equivalent capabilities to proliferate is not a planning horizon. It's a fire drill.

Whether you are a CISO, a security engineer, or a senior leader, this episode gives you the threat picture and a clear view of where the skills investment needs to go before the window closes.