AI
Anthropic's red team lets an unreleased Claude write five exploit chains
Anthropic's Frontier Red Team published five attack chains built by an unreleased internal research model against Chrome, Safari, an Android Pixel 9, Windows and the Linux kernel, including a one-click escape from the browser sandbox and a zero-click root chain that took 48 minutes to leak a kernel pointer. Every flaw went through coordinated disclosure and was patched before publication; Anthropic argues the point is that frontier models can now prove a bug by producing a working exploit.