Anthropic Vulnerability Ledger Shows Fix Delays and High Noise
Anthropic's Project Glasswing generated 37,966 vulnerability findings, but maintainers face high false positive rates and few fixes.
TL;DR
- Anthropic's Project Glasswing generated 37,966 vulnerability findings over seven months, but maintainers have marked only 217 as fixed [^1].
- A curl project review of 34 model reports revealed a 55.9% false positive rate, raising doubts about Anthropic's claimed efficiency [^1].
Background
Automated vulnerability scanning tools use software and artificial intelligence to find potential security flaws in codebase repositories. Security researchers often disclose these findings to open-source maintainers so developers can fix code before attackers exploit the flaws. However, AI model scanners can generate large volumes of unverified reports, forcing maintainers to spend time triaging automated submissions alongside regular development work.
What happened
Anthropic launched Project Glasswing on April 7th, claiming to identify thousands of high- and critical-severity vulnerabilities in open-source software [^1]. Seven months into the project, VulnCheck tracked 442 CVEs attributable to Anthropic researchers, representing about 0.5% of all 93,450 CVEs published this year [^1]. Out of those 442 CVEs, VulnCheck identified only two—CVE-2026-61500 and CVE-2026-26980—that have been exploited in the wild [^1].
Anthropic's public disclosure ledger expanded from 23,019 findings at launch in May to 37,966 findings by November 2nd [^1]. However, only 7,349 findings (19.4%) have been made public, and just 217 have been marked fixed [^1]. Between September 8th and November 2nd, Anthropic added 11,813 findings to its ledger while recording only 15 additional fixes [^1]. Meanwhile, 1,654 findings (4.4%) have been marked as duplicates and 419 have been withdrawn [^1]. Despite these duplicate and withdrawn reports, Anthropic's claimed true positive rate increased from 90.8% at launch to 94.2% [^1].
To address human triage bottlenecks, Anthropic announced its self-service OSS Scanner on October 8th, allowing open-source maintainers to receive unreviewed AI-generated reports directly [^1]. When curl maintainer Daniel Stenberg received 45 reports from Anthropic, 34 were published for public assessment [^1]. Maintainers determined that 19 reports were not vulnerabilities, six were low severity, and nine remained unassessed, yielding a 55.9% false positive rate so far [^1]. Earlier in May, five Mythos findings submitted to curl resulted in just one confirmed vulnerability [^1].
Why it matters
Shifting unvalidated AI scanner output onto open-source maintainers creates a substantial operational burden [^1]. When AI models generate thousands of automated reports, volunteer developers must pause core maintenance to review low-impact or invalid submissions [^1]. In curl's case, maintainer Daniel Stenberg had to pace the release of 45 reports across a week to soften the impact for his team [^1].
The gap between model severity ratings and maintainer reality highlights ongoing evaluation issues in AI vulnerability discovery [^1]. In September, Anthropic's Claude model rated 91.5% of ledger findings as critical or high severity, whereas maintainers rated only 51.3% at those levels [^1]. When automated findings stall without fixes, duplicate discoveries accumulate and drain open-source developer resources [^1].
Practical example
Say Mara is an open-source maintainer managing a popular C library.
Mara opts into Anthropic's self-service OSS Scanner to inspect her repository [^1]. Soon after, she receives a bulk delivery of dozens of model-generated reports that have not undergone human review or triage [^1].
Mara stops feature development to analyze each submission line by line. Similar to what happened with curl, where 19 out of 34 evaluated reports were not security flaws and six were low severity, Mara finds that more than half of the AI findings are false positives [^1]. A few reported issues turn out to be duplicates that other security tools already reported weeks earlier [^1].
Instead of getting actionable security fixes, Mara spends days writing detailed explanations to close invalid bug reports. Meanwhile, the unverified automated scanner reports add triage work to her project backlog while delivering minimal actionable security benefits [^1].