Subscribe
0%
Illustration for: Seven months of Anthropic's Glasswing: 442 CVEs, two exploited, and a fix count that depends on who you ask

Seven months of Anthropic's Glasswing: 442 CVEs, two exploited, and a fix count that depends on who you ask

VulnCheck's latest audit says Anthropic's AI bug hunt is burying maintainers in low-value reports; Anthropic's own dashboard claims a 94.2% true-positive rate and 629 upstream patches. We lined up both sets of numbers.

Listen to this article0:00 / 5:06

VulnCheck researcher Patrick Garrity published his fifth audit of Anthropic's Project Glasswing on November 9, and his verdict is blunt: seven months in, the AI-driven bug hunt is producing a flood of findings, a trickle of fixes, and a growing triage burden for open-source maintainers. He counts 442 CVEs credited to Anthropic, of which only two are known to be exploited.

Why it matters: Glasswing is the most visible test of whether frontier models can find real vulnerabilities at scale, and Anthropic has leaned on it heavily — including through its new OSS Scanner, which sends model-written reports directly to projects. Whether those reports save maintainers time or cost them time is now a live argument, and the two sides are reading the same data very differently.

The two sets of numbers

Both VulnCheck and Anthropic draw on Anthropic's public ledger, last updated November 2, 2026. They emphasize different lines.

Metric (to Nov 2) Figure Who highlights it
Candidate findings (every hypothesis Claude produced, pre-triage) 37,966 Both
Reviewed by external security firms 6,376 Anthropic
Confirmed valid by those firms 6,009 (94.2%) Anthropic
Reported to maintainers 5,087 (1,862 after outside triage; 3,225 sent direct) Anthropic
Patched upstream 629 Anthropic dashboard
Ledger entries marked "fixed" 217 VulnCheck
Marked duplicate / withdrawn ~1,654–1,670 / 419 VulnCheck
CVEs credited to Anthropic (VulnCheck tracker) 442 (~0.5% of 93,450 CVEs this year) VulnCheck
Of those, in VulnCheck's known-exploited list 2 VulnCheck

Sources: Anthropic's dashboard; VulnCheck. VulnCheck's own post lists the duplicate count as both 1,654 and 1,670.

The headline gap — 217 "fixed" versus 629 "patched upstream" — is the clearest example of why the two readings diverge. The 217 is a status count on individual ledger entries; the 629 is a separate dashboard metric for patches landed after a report was sent. Anthropic's documentation doesn't define "fixed" on the ledger, so neither number alone settles how many real bugs got closed.

VulnCheck's case

Garrity's sharpest data point is the trend since his September 8 snapshot. Candidate findings rose by about 11,800 to 37,966, but ledger "fixed" entries rose by only 15, from 202 to 217. Duplicates went from 2 to roughly 1,670, and withdrawals from 245 to 419.

He also questions the 94.2% figure. Anthropic's About page says its true-positive rate is the share of firm-reviewed findings confirmed as real — about 6,400 of nearly 38,000 candidates — and that it counts duplicates and "won't fix" findings as true positives. As CellCog notes, Anthropic doesn't publish an accuracy rate for the 3,225 reports it sent directly to maintainers.

The OSS Scanner, launched October 8, is his biggest concern. According to VulnCheck, Anthropic said the scanner has turned up over 29,000 candidates, about 6,000 of them manually reviewed, and that around 5,000 unverified reports had already gone to maintainers. VulnCheck quotes Anthropic describing these as fully model-generated, without human triage.

The curl test case

Garrity uses curl as a sample. Per VulnCheck, curl received 45 OSS Scanner reports; of the 34 published when he wrote, 19 were judged not to be vulnerabilities (55.9%), 6 were confirmed low severity, and 9 were pending. Anthropic had rated one report medium and the rest low. CellCog, citing curl's issue labels on November 11, counts 37 non-vulnerabilities, 7 confirmed lows and 1 pending; it also points out that VulnCheck's 55.9% treats real-but-non-security bugs as false positives.

It isn't all bad news for the scanner. curl maintainer Daniel Stenberg wrote in October that a HIGH-severity HTTP/2 use-after-free, CVE-2026-92392, was found and reported by an Anthropic researcher using the OSS Scanner and fixed in curl 8.23.0. In a separate post he said the junk AI reports of past years have largely stopped; the new problem is a high volume of higher-quality reports.

Anthropic's side

Anthropic hasn't published a point-by-point reply to VulnCheck, according to explainx.ai. Its November 2 dashboard notes say several totals fell because many reported findings turned out to be duplicates of other Anthropic findings and were removed from the counts.

Caveats on both sides: VulnCheck sells vulnerability intelligence; Anthropic's ledger updates in infrequent batches; and "only two exploited" lacks a baseline, since most CVEs are never exploited.

What to watch

  • Maintainers: if you opted into OSS Scanner bulk delivery, expect mostly low-severity or non-security reports. curl's approach, handling LOW and MEDIUM reports as public GitHub issues since October 26, is one way to keep the queue manageable.
  • Anthropic's next ledger update: watch whether "fixed" moves meaningfully, and whether Anthropic publishes accuracy for direct-to-maintainer reports.
  • Security teams: a credited CVE count isn't a risk measure. Prioritize by exploitation evidence, as you would any other source.

Sources

  1. VulnCheck — After 7-months, Glasswing Overwelms Maintainers with Underwhelming Findings vulncheck.com
  2. Anthropic — Coordinated vulnerability disclosure dashboard red.anthropic.com
  3. Anthropic — About this dashboard red.anthropic.com
  4. Daniel Stenberg — daniel.haxx.se blog (curl) daniel.haxx.se
  5. CellCog — Project Glasswing Results: VulnCheck vs Anthropic's Ledger cellcog.ai
  6. explainx.ai — Glasswing 37,966 Findings vs 217 Fixes explainx.ai

#Anthropic#Project Glasswing#VulnCheck#curl#open source#vulnerability disclosure#AI security