Patrick Garrity of VulnCheck published a seven-month review of Project Glasswing, Anthropic’s AI vulnerability-finding effort, on November 9, 2026. The review tracks 442 CVEs credited to Anthropic, about 63 a month, which Garrity puts at roughly 0.5% of the 93,450 CVEs published this year. VulnCheck lists two of those CVEs as known to be exploited.
Quick Take
- What: Anthropic’s Glasswing AI scanner produced 442 CVEs in seven months, with two known to be exploited in the wild (CVE-2026-61500, CVE-2026-26980).
- Who is affected: Open-source maintainers whose projects receive OSS Scanner reports carry the triage load, not Anthropic.
- Accuracy: Anthropic’s disclosure ledger records 217 findings as fixed while listing 1,654 duplicates and 419 withdrawn entries.
- curl example: curl received 45 Glasswing reports and published 34; VulnCheck counts 19 marked not a vulnerability, 6 confirmed as low severity, and 9 still awaiting assessment.
- What to do: Publish low and medium severity AI-generated findings as public issues so that the volume and accuracy of the reports stay visible.
In plain English
Anthropic built an AI tool called Glasswing that scans open-source code for security bugs. In seven months it found 442 problems that were serious enough to get official CVE tracking numbers. However, a security research firm called VulnCheck looked at the results and found that only two of those bugs are actually being exploited by attackers in the real world. Anthropic’s own records show that 1,654 of the reports were duplicates and 419 were withdrawn, while only 217 were actually fixed. The curl project received 45 reports and found that 19 were false alarms.
What you should do: If you maintain an open-source project and receive these AI-generated reports, you should publish them publicly so everyone can see how accurate they really are.
What happened
Patrick Garrity of VulnCheck published a seven-month review of Project Glasswing on November 9, 2026. The review tracks 442 CVEs credited to Anthropic, about 63 a month, which Garrity puts at roughly 0.5% of the 93,450 CVEs published this year.
VulnCheck lists two of those CVEs as known to be exploited: CVE-2026-61500 and CVE-2026-26980. The numbers do not line up with Anthropic’s own disclosure ledger, which records 217 findings as fixed while listing 1,654 duplicates and 419 withdrawn entries.
Glasswing findings reach maintainers through Anthropic’s self-service OSS Scanner, which Anthropic says sends model-generated reports without human review. Triage therefore lands on the receiving project rather than on Anthropic.
curl is the worked example in the review. The project received 45 reports and has published 34 of them. Of those, VulnCheck counts 19 marked not a vulnerability, 6 confirmed as low severity, and 9 still awaiting assessment. Garrity derives a 55.9% false positive rate from the published set, with 17.6% confirmed as genuine findings so far.
VulnCheck also compared Claude’s own severity ratings against the maintainers’ ratings of the same findings. Of the September ledger findings, 91.5% were rated critical or high by Claude, while maintainers rated only 51.3% of the same findings as critical or high.
What you should do
Garrity recommends that maintainers copy curl and publish low and medium severity AI-generated findings as public issues. Publishing them makes the real volume and accuracy of the reports visible to everyone, rather than leaving each project to absorb the triage privately.
Treat a model-generated report as unreviewed until a human confirms it. Anthropic sends OSS Scanner reports without human review, so the receiving project does the filtering.
What it means
The review shows that AI vulnerability scanners can produce a high volume of findings, but the triage cost falls on the receiving projects. The gap between Anthropic’s claimed true positive rate of 94.2% and the false positive rate VulnCheck derived from curl’s published reports suggests that the accuracy of AI-generated vulnerability reports is still a matter of debate.
The fact that only two of the 442 CVEs are known to be exploited in the wild, and that Anthropic’s own ledger lists far more duplicates and withdrawn entries than fixes, raises questions about the real-world impact of the effort.



