Anthropic Project Glasswing 7-Month Performance Review
Project Glasswing Performance Overview
Seven months after the launch of Project Glasswing, analysis of Anthropic’s vulnerability disclosure ledger indicates that the project has struggled to convert high-volume findings into actionable security fixes. While Anthropic has reported nearly 38,000 findings, only a small fraction have been verified and fixed, leading to concerns regarding the project's scalability and the accuracy of its reported true-positive rates.
Discrepancies in Disclosure and Remediation
Data tracked by VulnCheck reveals a significant disparity between the volume of findings generated by Anthropic’s models and the number of vulnerabilities actually addressed by maintainers. As of November 2026, Anthropic reported 37,966 findings, yet only 217 (approximately 0.6%) have been marked as fixed. Furthermore, the number of findings marked as duplicates (1,654) or withdrawn (419) has outpaced the number of confirmed fixes, suggesting that the triage process is failing to keep up with the volume of model-generated reports.
The Burden on Open Source Maintainers
Anthropic’s transition to an opt-in OSS Scanner has shifted the validation burden from Anthropic’s internal teams to individual open-source maintainers. Real-world testing by projects such as curl has highlighted the challenges of this model. In a recent batch of 45 reports submitted to curl, maintainers found that 19 were not vulnerabilities, 6 were low-severity issues, and 9 remained unassessed. This high false-positive rate—coupled with the manual effort required to triage these reports—has led to criticism that the initiative creates an unsustainable workload for maintainers without providing proportional security improvements.
Industry Perspectives and Counterpoints
Community response to these findings is divided. Critics argue that the project is overhyped and that the high volume of "slop" findings overwhelms security teams. Conversely, some maintainers report positive experiences, noting that the models can identify complex, valid vulnerabilities that human auditors might miss.
Industry observers also point out that the public ledger only represents a fraction of the project's total activity. Much of the high-value work occurs within private enterprise codebases where vulnerabilities are remediated internally without public CVE assignment. As one commenter noted, "Public datasets of CVEs being underwhelming doesn't mean much if all of the issues are in private company codebases that will never get a CVE number assigned."
Conclusion on AI-Driven Vulnerability Discovery
While AI models demonstrate clear utility in identifying potential security flaws, the current implementation of Project Glasswing highlights the limitations of automated discovery without robust, human-in-the-loop triage. The industry remains split on whether these tools represent a genuine shift in security posture or a marketing-led initiative that creates more noise than signal for the open-source ecosystem.
Sources
Related
- Dispatch
- Dispatch
- Dispatch
- Dispatch
- Dispatch