- VulnCheck researcher Patrick Garrity says he has tracked 442 CVEs tied to Anthropic's Project Glasswing over seven months, but says only two have been confirmed exploited in the wild.
- Anthropic's own public ledger lists 37,966 findings with just 217 marked fixed, according to Garrity's analysis—a remediation gap VulnCheck frames as a bottleneck, not a success.
- Curl maintainer Daniel Stenberg marked 19 of 34 published OSS Scanner reports as not actual vulnerabilities, a triage result that VulnCheck says undercuts Anthropic's claimed 94.2% true-positive rate.
What the Barn Doors Are Flapping About
Well, grab your sweet tea and settle in, because the security world is chewing on a seven-month report card for Anthropic's AI-powered vulnerability-hunting program, which the company describes as Project Glasswing. VulnCheck researcher Patrick Garrity published a blog post on November 9, 2026, laying out what he says his independent tracking of public CVE records shows, and the picture he paints is roughly as flattering as a mud-splattered pickup in a church parking lot.
According to Garrity's post, he has counted 442 CVEs attributable to Anthropic's initiative—about 0.5% of all CVEs published so far in 2026. Of those, Garrity says only two have been identified as exploited in the wild, a rate he pegs at approximately 0.45%. Separately, curl maintainer Daniel Stenberg publicly flagged that his project received a batch of 45 vulnerability reports from Anthropic's OSS Scanner on November 5, 2026, and the triage results on those reports have been, let's say, unimpressive.
What We Actually Know for Certain
Garrity's CVE count of 442 and his two-exploited-in-the-wild figure are drawn from public CVE attribution records, and those specific numbers have not been independently contested in the public record as of this writing. Anthropic's own public disclosure ledger—which the company itself maintains—showed 37,966 findings as of November 2, 2026, of which 7,349 were public and 217 were marked fixed, according to Garrity's analysis of that ledger. Those ledger figures reflect both sides counting different things, as Garrity himself acknowledges.
The curl data is the most concretely verifiable piece of this whole dustup. Daniel Stenberg, who maintains curl, stated publicly that of the 34 OSS Scanner reports published by the time of Garrity's post, 19 were marked 'not a vulnerability,' 6 were confirmed as low severity, and 9 remained under assessment. Stenberg's GitHub issue triage is publicly observable, and multiple outlets have quoted him directly, so that particular number ain't going anywhere. Separately, The Next Web reported—citing Garrity's broader work—that AI-discovered CVEs are arriving at roughly twice last year's rate, while the share that end up exploited has fallen to 1.4%.
What Nobody Has Pinned Down Yet
Anthropic claims its true-positive rate for Glasswing findings has climbed from 90.8% to 94.2%, but that figure is self-reported by the company with no independent audit behind it. Garrity publicly questions how duplicates and withdrawn findings are counted in that calculation, pointing to the curl triage numbers as evidence that the stated rate and the real-world rate may not rhyme. Anthropic, for its part, says most findings are still working through the disclosure pipeline, which would mean the 217-fixed figure is an early snapshot rather than a final grade.
Separately, Anthropic announced in October 2026—in its own announcement materials—that Project Glasswing had uncovered at least 129,000 verified software vulnerabilities between April and July 2026, with more than 33,000 rated critical or high severity. The Hacker News reported on that announcement, but the underlying verification of those figures rests entirely on Anthropic's own claim, and no independent source has fully replicated or contested Garrity's methodology end-to-end either. We got two silos of self-reporting staring each other down across a fence.
The Curl Situation Deserves Its Own Section
Stenberg's experience is worth looking at separately because it is the clearest window into what receiving an OSS Scanner dump actually feels like from the receiving end. According to Garrity's post, Stenberg got 45 reports in a single batch on November 5, 2026. Of the 34 that had been published by the time Garrity wrote his analysis, 19 were designated 'not a vulnerability.' That is a rejection rate of better than half on the published subset—a figure that sits uncomfortably next to Anthropic's claim of a 94.2% true-positive rate, even if the two data sets are measuring different things.
VulnCheck and Stenberg argue, according to Garrity's post, that Anthropic's OSS Scanner effectively offloads unreviewed, model-generated triage work onto volunteer open-source developers who have neither the time nor the obligation to serve as a free QA layer for a commercial AI product. Anthropic, the company says, frames the scanner as a free resource that empowers maintainers. Those two descriptions of the same situation are, as a certain neighbor of mine once said about two accounts of the same fistfight, technically about the same event.
Our Analysis: Volume Is the Hog, Impact Is the Ham
This is analysis, not settled reporting, but the gap between the numbers at the top of this story and the numbers at the bottom is wide enough to drive a tractor through. If Garrity's tracking is accurate, Anthropic's initiative is generating CVE volume at a scale the security world has not seen from a single source, while the slice of those findings that translates into patched code or stopped attacks remains extremely thin after seven months. That does not necessarily mean the program is failing—disclosure pipelines are genuinely slow, and downstream fixes take time—but it does mean the headline numbers Anthropic, the company, has promoted deserve to be held next to the remediation ledger before anyone hollers victory.
The OSS Scanner controversy is a preview of a broader question the industry has not answered: when AI systems generate vulnerability reports faster than human communities can triage them, who carries the load? Volunteer maintainers of open-source projects are not staffed for that job, and if the quality bar on inbound reports is lower than advertised—as the curl numbers suggest it may be, at least in that sample—the tool risks becoming a nuisance tax on the very ecosystem it claims to protect. Whether Anthropic adjusts its quality controls in response to this public pressure is something worth watching, though as of this writing the company has not announced any such change.
Who is doing the hollering
These links show where the chatter came from. A link is attribution, not our endorsement or independent confirmation.
- After 7-months, Glasswing Overwelms Maintainers with Underwhelming FindingsVulnCheck · specialist
- AI is finding twice as many software flaws. Almost none get exploited.The Next Web · top tier
- Anthropic Expands Claude Access for Vetted Cyber Teams as Glasswing Finds 129,000 FlawsThe Hacker News · specialist
- Project Glasswing Results: VulnCheck vs Anthropic's LedgerCellCog AI · specialist
- VulnCheck on Glasswing — 442 CVEs, 2 exploited, maintainers carry the triageAI TLDR · specialist
Last checked Nov 11, 2026, 9:07 AM EST. Talk Around Town: The core numbers—442 CVEs, 2 exploited, 217 fixes—come from a single VulnCheck analyst tracking public records; Anthropic disputes the framing and says most findings are still in the disclosure pipeline. Curl's triage is a real data point but covers just 45 reports from one project out of thousands.