Anthropic touts 94.2% true positives; curl rejected 19 of 34

The ledger’s own numbers

Anthropic’s Project Glasswing ledger listed 37,966 findings and 217 fixes as of its November 2 update, according to a seven-month tally VulnCheck published on November 10. If your project takes reports from Anthropic’s OSS Scanner, read the ledger’s 94.2% true positive rate as Anthropic’s own figure. The scanner’s reports are “fully model-generated, without human review or triage,” and in the one batch a maintainer has triaged in public, curl’s, more than half of the reports assessed so far were marked as not a vulnerability.

When the ledger launched in May, it showed 23,019 findings and a 90.8% true positive rate. By September 8 it held 26,153 findings and claimed 91.4%. It now claims 94.2%. Over the two months since September, Anthropic added 11,813 findings and only 15 fixes.

Of the 37,966 findings, 7,349 (19.4%) have been made public. Another 1,654 (4.4%) are marked as duplicates and 419 have been withdrawn, so duplicates outnumber fixes nearly eight to one. VulnCheck says duplicates went from two in September to more than 1,600 now, and says it predicted this: the longer findings sit, the more likely someone else finds and fixes the bug first. VulnCheck reads the rising true positive rate as likely a missing feedback loop, with duplicates and withdrawals not flowing back into what the ledger counts as true positives. It also reports a bug on the ledger that inflates the number of findings shown as disclosed.

What curl found

Daniel Stenberg, who maintains curl, announced on November 7 that the project had received a batch of 45 reports from Anthropic. Anthropic rated one of them medium severity and the rest low. Stenberg is moving the reports to GitHub at a deliberate pace “to soften the impact for the team” and plans to publish them all by the end of the week.

Of the 34 published so far, 19 are marked not a vulnerability, six are confirmed as low severity and nine still await assessment. VulnCheck puts the false positive rate at 55.9% so far, with 17.6% confirmed. Those rates can still move with the nine pending reports, and they come from one project. Still, curl has seen this before: in May, Stenberg wrote that five Mythos findings came down to a single confirmed vulnerability.

Severity tells a similar story at ledger scale. In September, VulnCheck found that Claude rated 91.5% of ledger findings critical or high, while maintainers rated 51.3% that way. Glasswing launched on April 7 on the promise of “thousands of additional high- and critical-severity vulnerabilities” headed for open source maintainers.

Why the triage moved to maintainers

Anthropic has acknowledged that human review did not keep pace. In its October 8 OSS Scanner launch post, it said it had found more than 29,000 candidate vulnerabilities, had manually reviewed about 6,000, and remained “bottlenecked on our human capacity to validate these findings.” By then it had already sent nearly 5,000 unverified reports to maintainers who had asked for everything in bulk.

OSS Scanner is opt-in. A project requests a scan and triages the findings itself, and Anthropic warns that “it is possible reports will be incorrect or invalid.” Projects that opt in therefore inherit the validation step that capped Anthropic’s own review. VulnCheck points to curl’s choice to publish the low and medium findings as public issues and encourages other maintainers to do the same.

CVEs and exploitation

VulnCheck builds its count from CVE records, GitHub advisories, Linux kernel commits and Anthropic’s ledger, plus manual audits of advisories. By that method it attributes 442 CVEs to Anthropic, about 0.5% of the 93,450 CVEs published this year, and more than twice the 217 fixes on Anthropic’s own ledger. In September the count stood at 225, with one exploited. Checked against VulnCheck KEV, VulnCheck’s catalogue of vulnerabilities seen exploited in the wild, two of the 442 now carry that flag: CVE-2026-61500 and CVE-2026-26980.