Anthropic is broadening access to its Claude AI models for vetted cybersecurity professionals, coinciding with a significant new milestone from Project Glasswing: the AI-assisted vulnerability research initiative has now identified more than 129,000 software flaws. The twin announcements signal that AI-powered security research is moving well beyond proof-of-concept territory and into large-scale, operational use.
Expanding the Circle of Vetted Access
The program, which requires organizations to pass a verification process before gaining access to Claude's more capable features for offensive and defensive security work, has been growing steadily since its launch. Anthropic's cyber verification program was designed specifically to prevent misuse while giving legitimate security researchers the tools they need to find vulnerabilities before malicious actors do. By adding more vetted teams, Anthropic is effectively widening the pool of professionals who can deploy Claude in higher-risk security contexts, with appropriate safeguards in place.
Key Facts
- Project Glasswing has identified over 129,000 software vulnerabilities to date
- Anthropic is expanding Claude access specifically for verified cybersecurity teams
- The verification program screens organizations before granting elevated access
- Glasswing targets open source software projects at scale
- The effort represents one of the largest AI-assisted vulnerability discovery initiatives on record
The scale of what Glasswing has accomplished deserves some context. What began as a focused experiment has evolved into one of the most productive automated vulnerability discovery efforts in the security community's history. Claude Mythos flagged 23,000 flaws across 1,000 open source projects at an earlier stage, and the total has grown dramatically since. Reaching 129,000 identified flaws means the pipeline is operating at a pace that human researchers working alone simply could not match.
The goal has always been to get ahead of attackers by finding vulnerabilities in widely used software before they can be exploited. AI allows us to do that at a scale that was previously unthinkable.Glasswing Research Team
What the Numbers Mean for the Security Community
Not every flagged issue is a critical zero-day, but the volume still matters. Even if a fraction of those 129,000 findings represent genuine, exploitable vulnerabilities in widely deployed open source code, the downstream impact on software security could be significant. Many of the projects targeted by Glasswing underpin web infrastructure, developer tooling, and enterprise software used by millions of organizations worldwide.
The expanded access program reflects a broader strategic bet by Anthropic that AI can be a net positive for cybersecurity when deployed responsibly. Rather than restricting Claude's use in security contexts altogether, the company has chosen a tiered model: general users get baseline capabilities, while verified professionals with clear, legitimate use cases get more. It is a pragmatic approach that acknowledges the dual-use nature of powerful AI tools without defaulting to blanket restrictions.
The trajectory of Glasswing also raises questions about what comes next. The project has grown from initial pilots to a sustained, high-volume operation. As Anthropic opens its latest Claude models to more cybersecurity firms, the expectation is that the pace of discovery will continue to increase. The security community will be watching closely to see how software maintainers respond to the flood of reports, and whether the disclosure pipeline can keep up with the volume of findings the AI is generating.
For now, the 129,000 figure is the headline number, but the more consequential development may be the infrastructure Anthropic is building around responsible AI use in security. That structure, combining model capability with an access control layer tied to real-world verification, could become a template for how the industry handles AI in other high-stakes domains.
