Claude Mythos Preview Uncovers 10k High-Risk Software Vulnerabilities (www.ibtimes.sg)

Anthropic's unreleased AI model, Claude Mythos Preview, has unveiled over 10,000 high- and critical-severity vulnerabilities in major software systems, marking a significant advancement in cybersecurity practices. This revelation comes from Project Glasswing, aimed at securing essential software against increasingly sophisticated AI-driven attacks. With participating organizations experiencing a tenfold increase in vulnerability detection, the need for rapid verification and patching has become more urgent than ever, fundamentally altering the software security landscape. Among notable findings, Cloudflare reported 2,000 bugs—including 400 critical vulnerabilities—attributable to Mythos Preview, which also outperformed human testers in accuracy. While over 23,000 vulnerabilities were identified across more than 1,000 open-source projects, the pace of discovery poses a challenge for cybersecurity teams, who struggle to address the overwhelming influx of issues, particularly in open-source software where some maintainers have requested a slowdown in disclosures. Despite its capabilities, Anthropic remains cautious about public deployment of Mythos, citing the need for better safeguards against potential misuse.