ResearchFeatured4 min read

Project Glasswing Finds 10,000+ Software Flaws in 30 Days

Anthropic's defensive security initiative used an unreleased Claude model to surface more than 10,000 high- or critical-severity vulnerabilities, including decades-old flaws in widely used software.

E
Editorial
May 31, 2026

Anthropic published its first progress update on Project Glasswing on May 22, 2026, reporting that its Claude Mythos Preview model and roughly 50 partner organizations identified more than 10,000 high- or critical-severity vulnerabilities in systemically important software during the program's first month.

Of the initial flagged issues, 6,202 were classified as high or critical across more than 1,000 open-source projects. Subsequent expert review confirmed 1,726 as valid true positives, of which 1,094 were rated high or critical. Notable findings included a 27-year-old remote crash flaw in OpenBSD, a 16-year-old flaw in FFmpeg, and CVE-2026-5194, a WolfSSL vulnerability scored CVSS 9.1.

Partner outcomes were significant. Cloudflare found 2,000 bugs in its own systems, 400 of them high or critical, while Mozilla fixed 271 Firefox vulnerabilities, a roughly tenfold increase over its experience with an earlier Claude model.

Anthropic is maintaining a 90-day coordinated disclosure policy, so specific details stay private during remediation. Analysts noted the update's central implication: the bottleneck in AI-assisted security has shifted from finding vulnerabilities to fixing them.

E
Editorial
May 31, 2026 · 4 min read
Back to News