AI Firm Anthropic Finds 10,000 Security Vulnerabilities in Critical Global Software

Artificial intelligence (AI) development company Anthropic has announced that it has discovered over 10,000 security vulnerabilities in critical software worldwide using its AI model 'Claude Mitos'. This achievement was made through Anthropic's 'Project Glasswing' cybersecurity initiative, launched last month.

Among the vulnerabilities found, 6,202 were classified as high-risk or critical flaws affecting over 1,000 open-source projects. Of these, 1,726 were confirmed as actual exploitable vulnerabilities, with 1,094 evaluated as high-risk or critical.

A representative example includes the discovery of the CVE-2026-5194 vulnerability in the encryption library WolfSSL.

Through this project, 97 vulnerabilities have been patched, and 88 security advisories have been issued. One participating bank contributed to preventing a $1.5 million wire transfer fraud by utilizing the AI model.

Anthropic highlighted the potential for AI security models and emphasized that companies must shorten patch cycles and strengthen network security settings.