Expanding Project Glasswing
Project Glasswing began with roughly 50 partners scanning codebases with Claude Mythos Preview, and they have already found over 10,000 high- or critical-severity security flaws. The program is now expanding to approximately 150 new organizations across more than 15 countries, covering critical infrastructure sectors like power, water, healthcare, communications, and hardware that were underrepresented in the initial cohort. Many new partners are vendors whose codebases are relied upon by governments and organizations worldwide, where a successful attack could affect more than 100 million people.
The article frames this expansion against an urgent timeline: within 6 to 12 months, many AI companies are expected to have Mythos-class models that could be released without safeguards against misuse. The bottleneck in cybersecurity has shifted from finding vulnerabilities to verifying, disclosing, and patching them at scale. Mythos Preview is already being used to write patches, perform pre-release checks, and simulate attacks for penetration testing. To accelerate this work, the team has released Claude Security, a product using public frontier models like Claude Opus 4.8 to scan codebases and suggest patches, alongside tools developed specifically for Project Glasswing partners.
The practical takeaway is that the cybersecurity landscape is about to shift dramatically as cheap, fast AI models with powerful cyber capabilities become widely available. The long-term goal is to give defenders a permanent advantage by safely providing broad access to better models, tools, and common infrastructure, while shifting support from vulnerability discovery toward disclosure, fixing, and deployment of patched software. Hundreds of thousands of organizations will likely need access to advanced cyber capabilities, but releasing Mythos-level capabilities in general access requires safeguards that do not yet exist. Project Glasswing serves as a template for how to respond when models cross important capability thresholds.
