Anthropic's Glasswing: 7 Months, Thin Results
Original: Tracking Anthropic-discovered vulnerabilities
Why This Matters
AI-driven vuln discovery is a major industry bet; actual remediation rates matter.
Seven months into Project Glasswing, VulnCheck has tracked 442 CVEs tied to Anthropic — just 0.5% of 2026's total — with only 2 exploited in the wild and 217 marked fixed out of 37,966 reported findings.
VulnCheck security researcher Patrick Garrity has been independently tracking Anthropic's AI-powered vulnerability discovery project, Glasswing, since its April 7 launch. The numbers paint a picture at odds with the original hype. Of 37,966 findings Anthropic has reported, only 7,349 (19.4%) are public and just 217 have been marked fixed. Since September, Anthropic added nearly 12,000 new findings while recording only 15 additional fixes — a widening gap between discovery and remediation that raises real questions about the project's operational impact.
Garrity's independent tracking repo has catalogued 442 CVEs attributable to Anthropic across CVE records, GHSAs, Linux kernel commits, and Anthropic's own ledger — more than twice what Anthropic's own disclosure ledger shows as fixed. That's an average of 63 CVEs per month, or 0.5% of the 93,450 CVEs published so far this year.
The severity question is equally pointed. Only 2 of the 442 CVEs have been confirmed exploited in the wild (0.45%), contradicting Glasswing's launch claim of discovering "thousands of high- and critical-severity vulnerabilities." Meanwhile, 1,654 findings (4.4%) are marked duplicates — nearly eight times the number fixed — and 419 have been withdrawn. Anthropic's self-reported true positive rate has climbed from 90.8% to 94.2%, a figure Garrity questions given how duplicates and withdrawals are accounted for. The curl project rejected 19 of 34 Anthropic-published reports as not being vulnerabilities at all, confirming only six as low severity.