Anthropic's AI bug hunter filed 442 CVEs, curl says most are false

Anthropic says its Project Glasswing security scanner catches real bugs 94.2 percent of the time. One of its heaviest-reported projects says most of what it files is not a bug. The maintainer-side numbers come from VulnCheck. Patrick Garrity went through Anthropic's own public ledger on 9 November. He writes that the pitch and the paperwork do not line up.

Anthropic launched Project Glasswing on 7 April 2026 and the free OSS Scanner on 8 October. Garrity's audit covers the public ledger as of 2 November. He counts 37,966 total findings and 7,349, or 19.4 percent, made public. Only 217 are marked fixed. 1,654 findings, 4.4 percent, are marked as duplicates, and 419 are withdrawn. 442 CVEs have been attributed to Anthropic in that time. That is 0.5 percent of the 93,450 CVEs published in 2026 so far.

The curl case

The curl project is the worked example in the post. Anthropic sent 45 reports to curl. 34 were published. Of those 34, curl marked 19 as not a vulnerability. Six were confirmed as low severity. Nine are still being checked. Garrity puts the false positive rate on curl at 55.9 percent. Anthropic says the scanner's true positive rate is 94.2 percent across the whole ledger, up from 90.8 percent at launch. The two numbers cannot both be true for curl without an explanation Anthropic has not given.

What happened in the wild

Only 2 of the 442 Anthropic-attributed CVEs have been exploited in the wild, by VulnCheck's tracking. Garrity names CVE-2026-61500 and CVE-2026-26980. He is careful to say that is not itself a mark against the scanner. Most CVEs are not exploited. The share is what matters for prioritisation, and two in 442 is low by any measure.

The volume problem for a maintainer

The gap between the Anthropic ledger and the curl assessment is a maintainer cost. That is the finding to sit with. A security report lands in a project's inbox whether it is real or not. Reading, reproducing and rejecting 19 reports in one quarter is unpaid work. The curl maintainers do it for free so a vendor can quote its precision number. Garrity notes that nearly 12,000 findings were added since September. Only 15 more fixes were recorded in the same period. Duplicates jumped from 2 to 1,670. The scanner is filing more reports every month, and the share of them that close as fixes is falling.

A team building a scanner that reads code with a model will take two things from this. First, the public precision number is only as good as the project that checks it, and curl is checking. Second, the work the model saves on the scanning side is work a maintainer has to spend sorting. An eval that scores the scanner has to count the reviewer time it uses. If it does not, the scanner improves in the wrong direction.

Source

Tracking Anthropic-discovered vulnerabilities, by Patrick Garrity at VulnCheck.