Anthropic Expands Project Glasswing to 150 Organizations in 15+ Countries

Anthropic has announced the expansion of Project Glasswing — a security program that uses powerful AI models to find vulnerabilities in critical software. Now approximately 150 new organizations in more than 15 countries are joining it.

Early Successes

Project Glasswing launched in April 2026 with 50 initial partners who gained access to Claude Mythos Preview. Within weeks, partners successfully used the model to scan their own codebases and discovered over 10,000 high and critical severity vulnerabilities. This demonstrated that AI models can significantly improve cybersecurity at the infrastructure level.

Expansion to Critical Sectors

New Project Glasswing partners operate in areas that were underrepresented in the initial group: energy, water supply, healthcare, telecommunications, and hardware manufacturing. Many of them are vendors — companies and non-profit organizations whose codebases are used by millions of other organizations worldwide, including governments. All 150 organizations are in more than 15 countries. Each underwent rigorous security screening before gaining access to Claude Mythos Preview. The key common characteristic: a successful cyberattack on their code would be catastrophic. For most partners, such an attack could affect over 100 million people and have serious consequences for both national and global security.

• Energy systems
• Water supply systems
• Healthcare organizations
• Telecommunications providers
• Equipment manufacturers

Preparing for the Era of AI-Powered Cyber Threats

Anthropíc emphasizes that cheap and powerful AI models with cyber capabilities will emerge very soon. The company expects that within 6-12 months, other AI firms will develop Mythos-level models and may release them without protective mechanisms. In such a reality, cyberattacks will occur more frequently and in unpredictable forms. Anthropic has defined a dual role for itself. First, to help the industry adapt by providing secure access to the best models, tools, and shared infrastructure. Second, to gradually transition from vulnerability discovery to disclosure, remediation, and patch deployment.

New Security Tools

The company recently released Claude Security — a public product based on Claude Opus 4.8 that scans codebases and suggests patches. It is also beginning to release tools developed for Project Glasswing partners that enable faster vulnerability discovery (distributed upon request among trusted security teams).

"Cybersecurity defenders must adapt to a reality where powerful AI

models are available to everyone."

What This Means

Project Glasswing is not simply a bug-finding program. It is an attempt to rethink the very model of cybersecurity in an era when AI models will become available to everyone. Anthropic is shifting work with critical software into a proactive defense mode, where vulnerabilities are discovered and remediated systematically, rather than reactively.