VulnCheck says Anthropic's Glasswing fixed few reported flaws

Anthropic’s Glasswing programme has yielded a vast number of vulnerability findings but delivered relatively few fixes, according to vulnerability researcher Patrick Garrity of VulnCheck. Garrity’s November 9 update on a VulnCheck blog grounds his assessment in publicly reported data from Anthropic, describing Glasswing as a mechanism to invite researchers to find and remediate flaws in Anthropic’s frontier AI models.

The latest CVE-related figures show that Glasswing surfaced 37,966 vulnerabilities, yet only 7,349 (19.4%) are recorded in Anthropic’s public disclosure ledger. Of these, 5,276 are validated as real unique vulnerabilities; 1,654 are duplicates and 419 withdrawn. Among the relevant vulnerabilities, about 12.9% (4866) have been reported to maintainers, while 193 remain in pre-disclosure and unreported to maintainers.

Garrity notes that 217 vulnerabilities—less than 0.5% of the total claimed through Glasswing—have been fixed. He further highlights that after September’s analysis, Anthropic added 4,613 findings but only 15 were marked as fixed in the same period, with many marked as duplicates.

Anthropic has said it launched an Open Source Security Scanner (OSS Scanner) to scale disclosures and reduce triage bottlenecks, enabling maintainers to request scans and triage themselves. The OSS Scanner follows a model inspired by Google’s OSS-Fuzz and has already delivered nearly 5,000 unverified vulnerability reports to maintainers upon request. Anthropic did not respond to Infosecurity’s request for comment by publication.