Claude Mythos Discovers Over 10,000 Critical Bugs in One Month

Anthropic has announced a significant breakthrough in cybersecurity with its latest AI model, Claude Mythos. Through Project Glasswing, an initiative providing limited access to the unreleased, highly capable AI to a select group of companies and governments, Mythos has uncovered over 10,000 high- or critical-severity vulnerabilities in widely used software within its inaugural month. This impressive performance highlights the model's advanced ability to identify complex security flaws that might otherwise go undetected for extended periods. According to Anthropic, partners involved in Project Glasswing have experienced a dramatic acceleration in their bug-finding efforts, with some reporting a more than tenfold increase in the rate at which they discover critical vulnerabilities. For instance, Cloudflare, a major internet hosting platform, utilized Mythos to find 2,000 bugs across its critical systems, with 400 classified as high- or critical-severity. Cloudflare's team even noted that Mythos Preview exhibited a lower false positive rate compared to human testers. The rapid and accurate identification of such a large volume of vulnerabilities by Claude Mythos indicates a potential transformation in the cybersecurity landscape. Anthropic suggests that the bottleneck in securing software is shifting from discovery to the subsequent processes of patching and responsible disclosure. This development underscores the immense potential of frontier AI models to bolster global cybersecurity defenses, even as the company acknowledges the powerful and potentially dangerous capabilities of Mythos, which led to its decision not to release it publicly.