Anthropic identified over 10,000 vulnerabilities using artificial intelligence

Anthropic has announced the expansion of its Project Glasswing program. This initiative focuses on identifying vulnerabilities in critical software using the specialized Claude Mythos Preview AI model. In the project's initial weeks, participants successfully identified over 10,000 high and critical-risk vulnerabilities. The number of organizations in the program will now increase from 50 to 200. This is reported by Ixbt.com .

Project Glasswing began as a closed collaboration between Anthropic, open-source software developers, and government entities. Participants gain access to a specially tuned Claude Mythos Preview model to find security flaws in large codebases. The new expansion wave will cover 150 additional organizations operating in the energy, healthcare, and communications sectors across more than 15 countries.

Anthropic experts believe that within the next 6–12 months, models at the Mythos level will become available to other developers. This could lead to a sharp increase in the number and complexity of cyberattacks. The company emphasizes that the challenge is shifting from finding vulnerabilities to remediating them rapidly. To this end, the Claude Security product was introduced, designed for automatic code analysis and patch preparation.

Currently, program participants are using the Claude Mythos Preview model not only to search for vulnerabilities but also to conduct security tests and modernize legacy software. Anthropic plans to make these capabilities available to all cybersecurity professionals in the future. As AI capabilities grow, the speed of adaptation by defenders becomes a key factor in digital infrastructure security.