Patrick Garrity of VulnCheck said in a Nov. 10 blog post that seven months after Anthropic launched Project Glasswing, he has independently tracked 442 CVEs attributable to Anthropic, while Anthropic's own disclosure ledger as of Nov. 2 listed 37,966 findings and 217 marked fixed¹.

Garrity said he began the open-source Anthropic Credited CVEs repository on April 17, using CVE records, GitHub Security Advisories, Linux kernel commits, Anthropic's ledger and manual advisory audits¹. He put the 442 CVEs at about 63 a month and about 0.5% of 93,450 CVEs published this year¹. He said that total is more than twice the 217 fixed items on Anthropic's public ledger¹.

Anthropic launched Glasswing on April 7 and said it had identified thousands of additional high- and critical-severity vulnerabilities it was disclosing to open-source maintainers, according to the post¹. Garrity compared the 442 CVEs with VulnCheck's known-exploited catalog and said two, CVE-2026-61500 and CVE-2026-26980, or 0.45%, had been identified as exploited in the wild¹.

He said the ledger first listed 23,019 findings and a 90.8% true positive rate in May¹. As of Nov. 2 he reported 7,349 findings (19.4%) made public, 217 fixed, 419 withdrawn and about 1,654 duplicates (4.4%), with Anthropic claiming a 94.2% true positive rate¹. His September-to-November table put duplicates at 1,670 on Nov. 2, up from 2 on Sept. 8¹.

From Sept. 8 to Nov. 2, Garrity counted 26,153 findings rising to 37,966 (plus 11,813), public items from 2,736 to 7,349, and fixes from 202 to 217 (plus 15)¹. He said withdrawn items rose from 245 to 419¹. He also said Anthropic's ledger appears to inflate the count of items marked disclosed when using a disclosed filter¹.

He cited Anthropic's Oct. 8 OSS Scanner launch post as saying the company had found more than 29,000 candidate vulnerabilities, had manually reviewed about 6,000, remained bottlenecked on human validation, and had sent nearly 5,000 unverified reports in bulk to maintainers who requested them¹. Anthropic described OSS Scanner reports as fully model-generated without human review and said reports may be incorrect or invalid, he wrote¹.

Garrity said curl maintainer Daniel Stenberg announced a batch of 45 Anthropic reports, one rated medium and the rest low, and was publishing them gradually on GitHub¹. Of 34 published so far, Garrity said 19 were marked not a vulnerability, six were confirmed low severity and nine awaited assessment, which he calculated as a 55.9% false-positive share so far and 17.6% confirmed, all low¹. He said curl in May had treated five Mythos findings as one confirmed vulnerability¹.

In September, Garrity wrote, Claude rated 91.5% of ledger findings critical or high while maintainers rated 51.3% that way¹. The post did not include a new on-record comment from Anthropic on the November figures¹.