Glasswing at seven months: 37,966 findings, 217 fixes, 1,654 duplicates
VulnCheck's seven-month audit of Anthropic Project Glasswing finds 442 attributed CVEs and a 2 November ledger of 37,966 findings against 217 fixes. Duplicates outnumber fixes about eight to one, and only two attributed CVEs are in VulnCheck KEV. The public ledger makes the backlog measurable; it does not establish attacker utility or a remediation rate.
Anthropic opened Project Glasswing on 7 April with a claim of thousands of additional high- and critical-severity flaws in open source, then stood up a disclosure ledger in May. A VulnCheck tracker started 17 April attributes 442 CVEs to Anthropic from CVE records, GitHub Security Advisories, Linux kernel commits, the ledger, and manual review. That is about 63 a month, 0.5% of 93,450 CVEs the post says were published this year. Two, CVE-2026-61500 and CVE-2026-26980, are in VulnCheck KEV.
On the 2 November ledger snapshot reported by VulnCheck, findings are 37,966, up from 23,019 in May. Public findings are 7,349, or 19.4%. Marked fixed: 217. Duplicates: 1,654, or 4.4%, roughly 7.6 times the fix count. Withdrawals: 419. Since September the same series records nearly 12,000 added findings and 15 added fixes. The claimed true-positive rate rose from 90.8% to 94.2%. Separately, curl maintainers rejected 19 of 34 published reports as vulnerabilities, confirmed six as low severity, and left nine unassessed.
The failure framing mixes units. 442 is an outside attribution of CVEs; 217 is a ledger state, not a concealed CVE total. A duplicate can be a true bug already reported, so 1,654 duplicates do not refute 94.2% until Anthropic publishes the numerator and the denominator. Two exploited CVEs out of 442 is 0.45%, but without the share of all 2026 CVEs in VulnCheck KEV that ratio has no baseline. No severity split is published, so the April high-and-critical claim is not confirmed or denied by the 442.
Disclosure clocks explain only the recent slice. Findings added since September can still be inside a 90-day window on 2 November. The May stock of 23,019 cannot. Discovery is scaling; patches are not. curl's 19 rejections in a set of 34 are one strict project's sample, not a global precision number, and they do show who pays for validation. The ledger's concrete gain is that a third party can count fixes, duplicates, and withdrawals at all. It does not yet show exploitation value or a maintainer-side path from report to patch.
Sources (3)
- [1]Tracking Anthropic-discovered vulnerabilities (VulnCheck, November update)(https://www.vulncheck.com/blog/anthropic-glasswing-november-update)
- [2]VulnCheck KEV(https://www.vulncheck.com/kev)
- [3]National Vulnerability Database(https://nvd.nist.gov/)