VulnCheck questions Anthropic's Glasswing: 37,966 candidate flaws, 217 fixes on record

Security firm VulnCheck published a November 11 update on Anthropic's Project Glasswing, and its central criticism is blunt: the flaws are being found, but not fixed. By early November the project had generated 37,966 candidate vulnerability discoveries, while the fix ledger VulnCheck tracks recorded only 217 repairs, according to a report by Chinese outlet QbitAI.

Glasswing launched in April with a large claim: use Claude at scale to hunt for security bugs in critical software and open-source projects, then work with technology companies and maintainers to patch them before attackers do. Anthropic pledged $100 million in model usage credits and $4 million in open-source security donations, said its security model had already surfaced thousands of high-severity and critical flaws, and drew in Apple, Google, Microsoft, AWS and Nvidia.

The gap opens between discovery and repair. Alongside the 217 fixes, VulnCheck's ledger lists 419 withdrawn records and 1,654 duplicates, and candidate findings jumped by nearly 12,000 in the past two months alone. The 442 CVEs VulnCheck attributes to Anthropic amount to just 0.5 percent of all publicly disclosed CVEs in 2026.

Anthropic counts differently. Its official numbers as of November 2 put 6,376 of the 37,966 candidates through external expert review, with 6,009 confirmed valid, 5,087 reported to software maintainers and 629 receiving upstream patches. Definitions explain part of the distance between 217 and 629, but both datasets point the same way: review and disclosure are not keeping pace with discovery.

Anthropic does not dispute the core problem. It has acknowledged that human verification capacity is a severe bottleneck limiting how fast findings can be reviewed and disclosed, and on October 8 it launched OSS Scanner, which lets open-source maintainers volunteer their code for free Claude scanning so reports go straight to them instead of waiting in Anthropic's review queue.

The cost lands on maintainers. curl's Daniel Stenberg received another batch of Anthropic reports in early November; he had already written on his blog that core maintainers were being overwhelmed, adding that "we aim to not drown." curl changed its vulnerability rules so low- and medium-severity reports are discussed publicly on GitHub, and of the 34 reports now public, 19 were judged not to be vulnerabilities and six were confirmed low severity.

That points to a harder problem. Generating a patch with an LLM is easy; getting an open-source project to adopt it is not. Repairs often add complexity, one fixed bug can breed ten new ones, and the review burden on maintainers does not shrink just because a model wrote the diff.

The report's own incentives are worth noting too: the gap VulnCheck criticizes, discovery without prioritisation, is precisely what its commercial vulnerability intelligence service sells. Even so, the structural issue stands. If AI raises the speed of finding bugs by an order of magnitude, the question is whether verification, disclosure and patching across the open-source ecosystem can scale with it.