Anthropic's latest model, Mythos, recently scored 94% on a benchmark for solving real software engineering problems—up from single digits just two and a half years ago. It also found a 27-year-old security flaw in OpenBSD that decades of human testing miss...