Anthropic’s flagship AI cybersecurity initiative, Project Glasswing, is facing fresh scrutiny after new research highlighted a widening gap between the number of software vulnerabilities being discovered and those being fixed.

According to a new analysis from VulnCheck, published on November 9, the initiative has generated tens of thousands of potential security findings since launching in April, but progress on remediation has been considerably slower.

Researcher Patrick Garrity examined Anthropic’s publicly available vulnerability disclosure records to assess whether the company’s AI-powered security efforts are delivering meaningful improvements.

As of November 2, Anthropic had recorded 37,966 potential vulnerabilities, but only 7,349 findings, or 19.4%, appeared in its public disclosure ledger.

Of those, Garrity identified 217 entries marked as fixed, alongside approximately 1,654 duplicates and 419 withdrawn findings.

The researcher also highlighted a growing disparity between discoveries and remediation. Between September 8 and November 2, the total number of findings increased by 11,813, while the number marked as fixed rose by just 15.

However, the figures require some qualification. Anthropic’s own disclosure dashboard reports that 5,087 vulnerabilities had been disclosed across 680 open-source projects as of November 2, with 629 patched upstream.

That broader figure differs from the number of entries Garrity identified as fixed in the public ledger, highlighting the importance of distinguishing between potential discoveries, disclosed vulnerabilities and confirmed remediation.

VulnCheck separately tracked 442 CVEs attributed to Anthropic researchers, equivalent to approximately 0.5% of all CVEs published during 2026. Just two of these had been identified as exploited in the wild.

AI Vulnerability Reports Raise Concerns

Beyond the number of fixes, the research raises questions about the practical burden AI-generated security findings are placing on software developers.

Anthropic claims a 94.2% true-positive rate for vulnerabilities assessed by external security research firms. However, this figure applies to a reviewed subset of findings rather than all 37,966 candidates, and can include genuine bugs subsequently identified as duplicates or considered unsuitable for fixing.

Garrity questioned whether the company’s reported accuracy adequately reflects the growing number of duplicate and withdrawn submissions.

The problem is particularly evident among open-source software maintainers, who increasingly face large volumes of AI-generated vulnerability reports requiring manual assessment.

Garrity pointed to the experience of curl, the widely used open-source data transfer tool, which recently received 45 vulnerability reports from Anthropic.

Of the 34 reports assessed in Garrity’s November 9 analysis, 19 had been classified as non-vulnerabilities, six were confirmed as low severity and nine remained under assessment.

Curl maintainer Daniel Stenberg has separately warned about the growing volume of vulnerability reports, which prompted the project to change how it handles low- and medium-severity submissions.

The findings nevertheless demonstrate both the opportunities and challenges of AI-assisted security research. Stenberg has also credited Anthropic with identifying a significant vulnerability in curl that was fixed in October.

A Growing Validation Bottleneck

Anthropic has acknowledged the difficulties associated with processing the volume of vulnerabilities its models can discover.

In October, the company announced an opt-in vulnerability scanning service for open-source maintainers, designed to make AI-generated findings available more quickly.

At the time, Anthropic said its models had identified more than 29,000 candidate vulnerabilities, but approximately 6,000 had undergone manual review.

Recommended reading

• Google’s Gemini AI Hacked Three Firms in Cyber Test
• Cyber Essentials Certifications Rise as SME Uptake Remains Limited
• Arnold Clark CISO Cuts Through the AI Cyber Hype
• Edinburgh-founded Quorum Cyber Appoints New CEO

The company acknowledged that it remained constrained by human capacity to validate findings, prompting efforts to scale its disclosure process.

Its new OSS Scanner allows participating maintainers to request scans and receive findings without the same level of independent human review, although this can transfer additional assessment responsibilities to already stretched development teams.

While Garrity acknowledged the value of AI in identifying software vulnerabilities, he argued that the evidence does not yet support the scale of Anthropic’s original claims.

The latest findings highlight a wider challenge emerging across the cybersecurity industry: as AI dramatically accelerates vulnerability discovery, organisations must develop equally effective ways to verify, prioritise and remediate the weaknesses it identifies.

Without those capabilities, the growing volume of security findings risks overwhelming the teams responsible for turning discoveries into meaningful protection.