TL;DR
Get privacy and security gear delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
VulnCheck says it has tracked 442 CVEs attributable to Anthropic since Project Glasswing began, while Anthropic’s ledger listed 217 findings as fixed as of Nov. 2. The tracker also reports thousands of findings that remain unpublished, duplicates or withdrawals; the figures do not establish how many are valid vulnerabilities or how many will be fixed.
VulnCheck says it has tracked 442 CVEs attributable to Anthropic in the seven months since Project Glasswing launched, while Anthropic’s public ledger listed 217 findings as fixed in its Nov. 2 update. The comparison adds detail to the project’s vulnerability-discovery effort, but the counts describe different measures and do not by themselves show how many findings are valid, exploitable or still awaiting repair.
VulnCheck’s report says its open-source tracker draws on CVE records, GitHub Security Advisories, Linux kernel commits and Anthropic’s ledger, with manual checks of advisories. It counted 442 CVEs it attributes to Anthropic, an average of about 63 a month. The report compares that tally with 93,450 CVEs published during the same year and says Anthropic-attributed CVEs amount to about 0.5% of that total.
Anthropic’s ledger, as reported by VulnCheck, showed 37,966 findings as of Nov. 2, up from 26,153 on Sept. 8. Of the November total, 7,349 were public and 217 were marked fixed. The ledger also listed 1,654 findings as duplicates and 419 as withdrawn. These are statuses reported in the ledger, not a separate independent validation of each finding.
VulnCheck says only two of its 442 tracked CVEs were listed in VulnCheck KEV as exploited in the wild: CVE-2026-61500 and CVE-2026-26980. The report also points to Anthropic’s claimed true-positive rate, which rose from 91.4% in September to 94.2% in the Nov. 2 snapshot. VulnCheck questions how duplicates and withdrawals are accounted for in that rate; the supplied figures do not resolve the methodology.
Discovery Counts Versus Repairs
The update highlights a distinction that matters to software maintainers and users: finding a possible vulnerability is not the same as confirming it, disclosing it, or fixing it. The gap between Anthropic’s reported findings and items marked public or fixed makes it difficult to infer from the headline count alone how much risk has been removed from software.
VulnCheck’s two exploitation cases offer a limited indicator of observed attacker use, not a measure of every finding’s potential severity. The report says those CVEs appeared in VulnCheck KEV, but the source material does not provide details about the affected products, exploitation timelines or the criteria used to connect them to Anthropic’s work.
The issue also affects how AI security tools are evaluated. A high claimed true-positive rate may sound reassuring, but readers need to know the denominator, review process and treatment of duplicates or withdrawn reports. Without those details, the percentage cannot on its own show how much useful, actionable security work the scanner produces.
Products Worth Considering
DJI Osmo Mobile 8 3-Axis Phone Gimbal Stabilizer
Jackery Explorer 300 Portable Power Station with SolarSaga 100W Solar Panel
Apple Watch SE 3 GPS 40mm Smartwatch
As an affiliate, we earn on qualifying purchases.
Glasswing’s Review Bottleneck
Anthropic launched Project Glasswing on April 7, saying it had identified thousands of additional high- and critical-severity vulnerabilities and was working on responsible disclosure. VulnCheck began its own tracking on April 17, before Anthropic created a public disclosure ledger. Its report says Anthropic updates that ledger in bulk rather than continuously.
Anthropic described a human-review constraint in an Oct. 8 post launching its OSS Scanner. According to the account quoted by VulnCheck, the company had found more than 29,000 candidate vulnerabilities but manually reviewed about 6,000, and said it was “bottlenecked on our human capacity to validate these findings.” Anthropic also said it had sent nearly 5,000 unverified reports in bulk to maintainers who requested them.
The OSS Scanner lets participating projects request scans, shifting part of the assessment work to maintainers. Anthropic describes those reports as model-generated without human review or triage and warns that they may be incorrect or invalid. That caveat is important when interpreting both the volume of candidate findings and the amount of work needed to establish whether a report warrants a fix.
““bottlenecked on our human capacity to validate these findings.””
— Anthropic, in its Oct. 8 OSS Scanner launch post, as quoted by VulnCheck
How Ledger Statuses Are Counted
The figures come from different sources and track different things: VulnCheck’s 442 is a count of CVEs it attributes to Anthropic, while Anthropic’s ledger totals findings across its project. The source does not establish that the two sets match one-for-one. Nor does the material explain the full method behind Anthropic’s true-positive rate, including how duplicates, withdrawals and findings still under review affect the calculation.
VulnCheck also says a ledger display bug appears to inflate the number of findings marked disclosed, based on its own inspection of the disclosed filter. The supplied report does not include a response from Anthropic confirming that issue or explaining its scope. The status counts should therefore be read as the report’s account of the ledger, not as settled evidence of the bug’s cause or effect.
For the two CVEs identified in VulnCheck KEV, the report provides no case-by-case information on severity, affected versions, patch availability or exploitation activity. It is also unclear how many of the unpublished findings will ultimately be confirmed, disclosed, fixed, withdrawn or marked duplicate.
Further Triage and Ledger Updates
The next developments to watch are new ledger updates, maintainer assessments and documented fixes. Anthropic’s OSS Scanner gives participating maintainers reports to review, but the source material does not specify a timetable for resolving the existing backlog or for updating the ledger more frequently.
VulnCheck says it will continue tracking CVEs it attributes to Anthropic and comparing them with exploitation records. A clearer picture will depend on additional disclosures and on more information about how Anthropic calculates its true-positive rate and assigns statuses. Until then, the published totals show the scale of reported activity and the limited number marked fixed, but not the final validity or remediation outcome of the full set of findings.
Key Questions
What did VulnCheck report?
VulnCheck says its tracker has identified 442 CVEs attributable to Anthropic since Project Glasswing began. It compared that count with Anthropic’s ledger, which listed 37,966 findings and 217 marked fixed as of Nov. 2.
Are all 37,966 ledger findings confirmed vulnerabilities?
The source does not establish that. Anthropic described many items as candidate findings, and it warns that OSS Scanner reports are model-generated and may be incorrect or invalid. The ledger’s total should not be treated as a count of independently confirmed vulnerabilities.
How many tracked CVEs were exploited in the wild?
VulnCheck says two of the 442 CVEs appeared in VulnCheck KEV as exploited in the wild. The report names CVE-2026-61500 and CVE-2026-26980; it does not provide further case details in the supplied material.
Why does VulnCheck question the true-positive rate?
Anthropic’s claimed rate rose from 91.4% in September to 94.2% in the Nov. 2 ledger snapshot. VulnCheck questions how the rate accounts for duplicates and withdrawn findings, but the supplied report does not provide enough information to independently determine the rate’s calculation.
What happens after an OSS Scanner report is sent?
Anthropic says the reports are sent without human review or triage, so maintainers must assess them. The report does not specify how quickly findings will be reviewed, disclosed or fixed.
Source: vulncheck.com
