Anthropic Project Glasswing: Mythos Preview flagged over 10,000 security flaws across critical systems

In a groundbreaking development, Anthropic has unveiled the initial results from its Claude Mythos preview, part of the ambitious Project Glasswing initiative. The AI-driven model has astonishingly identified more than 10,000 critical and high-severity vulnerabilities across widely-used software systems within just a month of its release. The Mythos Preview was shared with approximately 50 select partners, providing a glimpse into the rapid capabilities of the AI model to verify, disclose, and patch significant numbers of vulnerabilities. According to a detailed blog post by Anthropic, these partners collectively reported hundreds of critical vulnerabilities in their systems, with Claude Mythos independently identifying around 6,202 critical flaws out of a total of 23,019 flagged issues. Notably, of the vulnerabilities that were examined by security firms, a staggering 90.6% were confirmed as genuine security threats, with 62.4% classified as high severity. The potential impact of Mythos Preview is significant, as it could account for nearly 3,900 confirmed high- or critical-severity vulnerabilities in open-source software, in addition to those reported by Project Glasswing partners. Among the key findings, Cloudflare reported approximately 2,000 bugs in its internal software, including around 400 critical vulnerabilities. Mozilla also benefited from the testing of Mythos, discovering and rectifying over 271 vulnerabilities in Firefox 150, a tenfold increase compared to its previous version, Firefox 148, which utilized Claude Opus 4.6. While Anthropic acknowledges the immense potential of AI models like Mythos in enhancing cybersecurity measures, the company is exercising caution regarding a public release of the technology. They caution that such powerful tools could be misused if made widely available. Anthropic has indicated plans to develop and release additional cybersecurity models akin to Mythos Preview, emphasizing the urgent need for the software industry to address the growing volume of findings these models generate.