Claude Mythos found a 27-year-old bug in OpenBSD's TCP stack. Two packets. Any host crashes. Anthropic will not release the model publicly.
On Mozilla's Firefox JS engine, Mythos generated working exploits in 72% of trials. Opus 4.6 hit 14%. Sonnet 4.6 hit 4%. One model generation apart.
Project Glasswing restricts Mythos to a select group of organizations: AWS, Apple, Google, Microsoft, CrowdStrike, the Linux Foundation.
Defenders get the same tool attackers will eventually build.
The vulnerabilities are confirmed, not noise. OpenBSD maintainers shipped the patch. CrowdStrike and Palo Alto Networks are validating findings through Glasswing. These are shipping as fixes, not sitting in a backlog.
I run production serverless, npm audit on every deploy across 30+ stacks. On a daily basis, it catches known CVEs but not zero-days. When this capability ships broadly, the volume of zero-days hitting npm dependencies will break every patch pipeline still built for human speed.
The attack surface is not the bug. It is the gap between discovery and production patch.
If your patch window is still measured in weeks, you're defending against last year's frontier.