Key points
- VulnCheck tracked 442 CVEs attributable to Anthropic, with just two confirmed as exploited in the wild.
- Only 19.4% of Anthropic's reported findings have been made public, and just 217 have been marked fixed.
- The OSS Scanner shifts validation to maintainers, with curl rejecting 19 of 34 published reports as vulnerabilities.
The evidence
According to VulnCheck, seven months into the project, the company has tracked 442 CVEs attributable to Anthropic, which represents roughly 0.5% of all CVEs published this year.reported Of these, only two have been identified as exploited in the wild.The data shows that while Anthropic’s AI tool finds many potential vulnerabilities, very few are actually fixed by developers.The low remediation rate and high duplicate volume suggest that AI-generated vulnerability reports currently impose significant triage costs on open source maintainers. The broader ledger shows 37,966 reported findings from Anthropic, but only 7,349 (19.4%) have been made public.reported
Evidence: reported
After 7 months, Glasswing overwelms maintainers with underwhelming findingsaifleet-hn-front
Evidence: reported
After 7 months, Glasswing overwelms maintainers with underwhelming findingsaifleet-hn-front
Since September, the gap has widened. VulnCheck reports that Anthropic added nearly 12,000 findings while recording just 15 additional fixes.reported This contrasts with the initial launch claims on 7 April, where Anthropic stated it had identified thousands of high- and critical-severity vulnerabilities. The current data shows only 217 findings marked as fixed overall.reported
Evidence: reported
After 7 months, Glasswing overwelms maintainers with underwhelming findingsaifleet-hn-front
Evidence: reported
After 7 months, Glasswing overwelms maintainers with underwhelming findingsaifleet-hn-front
The decision
VulnCheck notes that roughly 1,654 findings (4.4%) were marked as duplicates, which is nearly eight times the number of fixes.reported Another 419 findings were withdrawn.reported These categories suggest that a significant portion of the initial reports may not represent unique or actionable security issues.reported
Evidence: reported
After 7 months, Glasswing overwelms maintainers with underwhelming findingsaifleet-hn-front
Evidence: reported
After 7 months, Glasswing overwelms maintainers with underwhelming findingsaifleet-hn-front
Anthropic’s claimed true positive rate has risen from 90.8% to 94.2%. However, VulnCheck questions how this metric accounts for the high volume of duplicates and withdrawn reports.reported
Evidence: reported
After 7 months, Glasswing overwelms maintainers with underwhelming findingsaifleet-hn-front
What to make of it
The OSS Scanner appears to shift the burden of validation onto open source maintainers. VulnCheck highlights that the curl project rejected 19 of 34 published reports from Anthropic as not being vulnerabilities.reported Six were confirmed as low severity, and nine remain awaiting assessment.
Evidence: reported
After 7 months, Glasswing overwelms maintainers with underwhelming findingsaifleet-hn-front
The data implies that the current model of AI vulnerability discovery may overwhelm maintainers without providing proportional security benefits. With only 15 fixes recorded in the most recent period despite 12,000 new findings, the operational cost of validation may outweigh the security gain for many projects.reported
Evidence: reported
After 7 months, Glasswing overwelms maintainers with underwhelming findingsaifleet-hn-front
In practice
Access to Anthropic’s disclosure ledger is limited to infrequent bulk updates, meaning real-time visibility into new findings is not available. Availability of fixes remains low, with only 217 of 37,966 reported issues resolved to date.reported
Why it matters
- Open source maintainers face increased operational overhead from triaging a high volume of AI-generated vulnerability reports, many of which are duplicates or non-issues.
- The low fix rate (217 out of 37,966) suggests that AI vulnerability discovery tools may not significantly improve the security posture of open source projects in the short term.
- The shift of validation work to maintainers, as seen with curl, indicates that AI tools are not replacing human expertise but are adding to the workload without clear efficiency gains.
- The discrepancy between Anthropic’s claimed true positive rate and the high number of duplicates/withdrawals raises questions about the reliability of AI-generated security metrics.
Topics: Anthropic, vulnerability management, open source security, AI tools, VulnCheck
Updated
The strongest doubt
Counterpoint from the sources
AI remains an effective tool for finding vulnerabilities, and the volume of findings may improve over time as the technology matures.
Numbers in context
Circle area is proportional to the magnitude; labels retain signs. Figures may describe different measures. Read their contexts above.Ask the article
Local keyword search of this article; no generated answers.Did this article give you the evidence you needed?
Your choice stays on this device. No public tally.Questions readers ask
How many CVEs has Anthropic been credited with so far?
VulnCheck tracked 442 CVEs attributable to Anthropic in the first seven months of Project Glasswing.
How did the curl project respond to Anthropic’s reports?
Curl rejected 19 of 34 published reports as vulnerabilities, confirmed six as low severity, and left nine awaiting assessment.
Connected stories
Reading receipt
- VulnCheck tracked 442 CVEs attributable to Anthropic, with just two confirmed as exploited in the wild.
- Only 19.4% of Anthropic's reported findings have been made public, and just 217 have been marked fixed.
- The OSS Scanner shifts validation to maintainers, with curl rejecting 19 of 34 published reports as vulnerabilities.
Still open: How does Anthropic account for duplicates and withdrawn findings in its claimed true positive rate?
Still open: Will Anthropic adjust its reporting methodology to reduce the volume of low-value findings sent to maintainers?