VulnCheck: Project Glasswing has 37,966 candidate vulnerabilities, 217 marked fixed in public records

VulnCheck reports that Anthropic's Project Glasswing has produced 37,966 candidate vulnerabilities in seven months, but only 217 are marked as fixed in the public disclosure ledger. The cybersecurity firm published the update on November 11, according to Chinese-language outlet QbitAI (量子位), which reported on it November 12.

What VulnCheck found

Project Glasswing launched in April. It uses Claude to find vulnerabilities in critical software and open source projects, working with tech companies and maintainers to fix them before attackers exploit them. At launch, Anthropic announced model usage credits and open source security donations. It also said its model Claude Mythos Preview had found thousands of high-severity and critical vulnerabilities. Apple, Google, Microsoft, AWS and Nvidia joined the program.

VulnCheck's tracking, as of early November, shows:

• 37,966 candidate findings from Anthropic
• 217 marked as fixed in the disclosure ledger
• 419 withdrawn entries
• 1,654 duplicates
• 442 CVEs attributed to Anthropic, which VulnCheck says is 0.5% of all public CVEs in 2026

VulnCheck says the gap is widening. Over the most recent two months, the candidate count rose sharply while fixed records increased by only 15.

Anthropic's figures

Anthropic's own disclosure, as of November 2, uses a different methodology. Of the 37,966 candidates, 6,376 have been reviewed by external security experts, and 6,009 of those were found valid. Of the valid findings, 5,087 have been reported to software maintainers, and 629 have received upstream patches.

Anthropic acknowledges that human validation capacity is a severe bottleneck that limits review and disclosure speed. On October 8, it launched OSS Scanner. Open source maintainers can sign up to have Claude scan their code for free, and the resulting reports go directly to maintainers without waiting for Anthropic's manual review.

The maintainer burden

VulnCheck cites curl. In early November, maintainer Daniel Stenberg received 45 vulnerability reports from Anthropic. He had previously written that core maintainers were overwhelmed by incoming reports: "We aim to not drown." He changed curl's handling rules so that low and medium severity issues are discussed publicly on GitHub, letting more community developers help review.

Of curl's 34 publicly disclosed reports, 19 were judged not to be security vulnerabilities and 6 were confirmed as low severity.

In September, Claude found a use-after-free memory error in curl. Under certain conditions, a malicious server could use it for remote code execution. Stenberg called it one of curl's most serious vulnerabilities in years.

Why patches lag

Anthropic has had Claude generate candidate patches. QbitAI cites developer complaints that the hard part is getting open source projects to adopt them, and that reviewing an AI patch can take nearly as much work as writing one. Developers also say AI fixes can add complexity and introduce new bugs.

QbitAI notes that VulnCheck sells vulnerability intelligence and prioritization services, which align with the problems its report criticizes.

What this means

The two datasets measure different stages of the pipeline. VulnCheck counts fixes marked in public disclosure records (217), while Anthropic counts valid findings that have received upstream patches (629). Both show far fewer fixes than candidates.

The constraint is human capacity at review and maintenance, not discovery. Anthropic's 6,376 externally reviewed findings are a small share of 37,966 candidates. The curl data shows that candidate volume does not translate directly into confirmed security issues: 19 of 34 public reports were not vulnerabilities. It also shows that individual findings can matter, as the September use-after-free case did.

OSS Scanner changes who does the triage, moving it to maintainers. Whether that raises the fix count depends on maintainer capacity, which is the pressure Stenberg describes.