Anthropic Expands Claude Access for Vetted Cyber Teams, Project Glasswing Identifies 129,000 Vulnerabilities

Anthropic on Tuesday said it's expanding a program that allows vetted cybersecurity professionals to test its advanced artificial intelligence (AI) models with reduced safeguards and blocking classifiers, as the company claimed its Project Glasswing initiative uncovered at least 129,000 verified software vulnerabilities between April and July 2026.

The expansion of the vetted access program reflects Anthropic's effort to balance AI safety with the need for deeper security research. By reducing automated blocking classifiers, the company intends to enable cybersecurity professionals to probe model behaviors and integrated systems more effectively. The company stated that the Project Glasswing initiative, running from April through July 2026, identified a substantial volume of vulnerabilities, underscoring the prevalence of software flaws in systems interacting with AI models.

Anthropic emphasized that the program is designed for responsible disclosure and vulnerability discovery. No active exploitation of the identified vulnerabilities has been reported. The company has not released a software patch, as the initiative focuses on identification and responsible reporting rather than immediate remediation. Organizations using Claude models are advised to review AI output handling and integrate any findings into their own patch management and security testing cycles.

The announcement comes amid broader industry discussion about the role of AI in security research and the challenges of securing increasingly complex software supply chains. The 129,000 identified vulnerabilities, while significant, reflect the scope of issues found during the specified testing window and are subject to independent verification.