Anthropic unveils new powerful AI that finds software flaws, but says it's too dangerous to release

Anthropic announced Project Glasswing.

It is providing Claude Mythos Preview—a frontier model—to selected major firms and more than 40 organizations for targeted scans.

Partners include Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, the Linux Foundation, Microsoft, Nvidia, and Palo Alto Networks.

The model has uncovered thousands of previously unknown vulnerabilities, including high-severity zero-day flaws.

Examples: an autonomously discovered 27-year-old OpenBSD remote-crash bug; a 16-year-old FFmpeg bug in a line run five million times; and Linux kernel vulnerability chains that escalate to full system exploitation.