Anthropic’s Project Glasswing initiative has identified over 10,000 software vulnerabilities using the Claude Mythos Preview AI model, leading to 97 upstream patches and 88 security advisories issued globally.

Key Points

• Project Glasswing provides 50 partners with early access to the Claude Mythos Preview model to autonomously detect critical software flaws.
• The initiative has confirmed 1,726 valid vulnerabilities, with 1,094 classified as high- or critical-severity across more than 1,000 open-source projects.
• A critical flaw in WolfSSL, identified as CVE-2026-5194 with a 9.1 CVSS score, was among the vulnerabilities discovered and addressed.
• The AI model successfully prevented a $1.5 million fraudulent wire transfer for a partner bank by detecting unauthorized account activity.
• Anthropic launched a Cyber Verification Program to allow security professionals to conduct red teaming and vulnerability research without standard AI guardrails.

Why it Matters

The rapid discovery of vulnerabilities by AI models creates an urgent need for software vendors to accelerate their patch cycles and strengthen network defenses. As these advanced tools become more prevalent, organizations must prioritize faster remediation to stay ahead of potential exploitation by malicious actors.