The AI company’s new Claude Mythos Preview model is “in a different league” from its predecessor in identifying and exploiting cybersecurity flaws, Anthropic said. Mythos is so good at hacking, according to its maker, that Anthropic won’t release it widely, but has launched an effort it’s calling Project Glasswing, inviting dozens of organizations that make and maintain critical software systems to use Mythos to find and patch dangerous bugs. (The Logic)

Talking point: Major participants in Glasswing include Microsoft, Apple, Amazon, Nvidia and Cisco. Mythos is good at finding subtle defects and ones that are only dangerous if several are strung together, Anthropic security researchers wrote in a technical description of their work. It identified a 27-year-old security hole in OpenBSD (an open-source operating system meant to be extra safe) and a 16-year-old vulnerability in FFmpeg (a very common set of tools for internet video and audio).