Project Glasswing is Anthropic's flagship effort to find vulnerabilities in open-source software with its AI models; it launched on 7 April 2026 with a promise of "thousands" of high- and critical-severity findings for maintainers. On 9 November, the security firm VulnCheck published its seven-month review, and it is far less flattering.
The numbers, according to VulnCheck: Anthropic's public ledger listed 37,966 findings as of 2 November, of which 217 were marked fixed, 419 withdrawn and about 1,654 duplicates. Since 8 September, findings rose by 11,813 while fixes rose by only 15. VulnCheck counts 442 CVEs attributed to Anthropic this year — about 0.5% of the 93,450 CVEs published — and only 2 of them are known to be exploited in the wild. It also says Claude rated about 91.5% of findings critical or high, against 51.3% when maintainers rated them, and that a likely bug inflates the ledger's "disclosed" count.
Anthropic's own dashboard counts differently. As CellCog summarises it, 6,376 candidates were reviewed by outside firms and 6,009 confirmed (the 94.2% "true positive rate" applies to that subset only); 5,087 were reported to maintainers, 3,225 of them sent without outside review; 629 were patched upstream; and 733 advisories got identifiers (230 CVEs and 503 GitHub advisories). The two sources' fix and CVE figures don't match, and Anthropic has not publicly answered VulnCheck.
What maintainers see: Anthropic's self-service OSS Scanner, launched on 8 October, sends reports that are "fully model-generated, without human review or triage", and Anthropic said it was "bottlenecked on our human capacity to validate these findings". curl received 45 reports on 5 November, one rated medium and the rest low. On 10 November its lead developer Daniel Stenberg wrote: "out of Anthropic's latest 45 findings in #curl, we are at 7 confirmed LOW." VulnCheck's conclusion: AI is a useful bug-finding tool, but the triage work is shifting onto the volunteers Glasswing was meant to help.
