Project Glasswing in numbers: what the VulnCheck audit found

Published 2026-11-11 · Last updated 2026-11-11

Short answer

Project Glasswing is Anthropic's program for finding vulnerabilities in critical software with Claude Mythos Preview. On November 2, 2026, Anthropic's dashboard listed 37,966 candidate findings and a 94.2% true positive rate. VulnCheck counts 442 CVEs it attributes to Anthropic, 2 of them exploited in the wild.

Every number below comes from Anthropic's or VulnCheck's own pages and links to the page it came from. All pages were checked on 2026-11-11. Where a source gives no number, the table says "Not stated".

What Project Glasswing is, per Anthropic

Anthropic announced Project Glasswing on April 7, 2026 as "a new initiative that brings together Amazon Web Services, Anthropic, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks in an effort to secure the world's most critical software" (announcement). The partners use Claude Mythos Preview, which the announcement calls "a general-purpose, unreleased frontier model", for defensive security work. Anthropic extended access to "over 40 additional organizations" and committed "up to $100M in usage credits for Mythos Preview" plus "$4M in direct donations to open-source security organizations" (announcement).

Anthropic also scans open-source projects itself and tracks the results on a public coordinated vulnerability disclosure (CVD) dashboard. Its process, as the dashboard describes it: Claude models produce candidate findings, one of six external security research firms reproduces and rates them, and confirmed bugs are reported to maintainers. Some findings go to maintainers directly, "without the same independent check", including when maintainers ask for un-triaged findings (CVD dashboard). The dashboard was last updated November 2, 2026 at 19:30 UTC. It keeps dated snapshots of earlier versions on its archive page.

On October 8, 2026 Anthropic launched OSS Scanner, an opt-in scanner for open-source projects. Anthropic says its reports "will be fully model-generated, without human review or triage" and that "it is possible reports will be incorrect or invalid" (OSS Scanner post).

What VulnCheck measured, and how, per VulnCheck

VulnCheck's post is by Patrick Garrity and dated November 9, 2026. It builds on four earlier VulnCheck posts. VulnCheck is a vulnerability intelligence company. The post reports four kinds of measurement:

• CVE tracking. Since April 17, 2026 it has kept an open-source repo, Anthropic Credited CVEs, that "tracks every vulnerability attributable to Anthropic using CVE records, GHSAs, Linux kernel commits, and Anthropic's own ledger", with manual audits of advisories (VulnCheck).
• Exploitation. It checks those CVEs against VulnCheck KEV, its own list of vulnerabilities exploited in the wild (VulnCheck).
• Ledger comparison. It compares counts from Anthropic's ledger in its September 8 post with the November 2 update (VulnCheck).
• curl reports. It counts the maintainer verdicts on the OSS Scanner reports that the curl project has published so far (VulnCheck).

Side by side

Anthropic's figures are as of November 2, 2026. VulnCheck's are as of its November 9, 2026 post.

Where the methods differ

Several rows above share a name but not a definition. Each source states its own.

True positive rate

Anthropic defines it as "The share of firm-reviewed findings confirmed as real, including duplicates and 'won't fix' findings" (About page). The dashboard adds that "the number of 'true positives' in the dashboard above should only be taken as one proxy for impact. Another, more reliable one is the number of patches created" (CVD dashboard).

VulnCheck's curl figure measures something else. It is the maintainer's verdict on OSS Scanner reports, which Anthropic says are not reviewed by humans (OSS Scanner post). VulnCheck puts the curl "false positive rate at 55.9% so far", with 17.6% confirmed, all low severity (VulnCheck). Anthropic's own check of an early scanner version had penetration testers review 97 critical and high findings across 48 projects. 85 (88%) met the bar for its CVD process (OSS Scanner post). These are three different reviewers looking at three different sets of reports.

Fixed

The dashboard overview counts 629 findings "Patched upstream", defined as "A patch has landed after the report was sent" (CVD dashboard, About page). The ledger's status filter shows 217 entries as "fixed" (ledger). VulnCheck uses 217 (VulnCheck). The pages we read do not explain the difference between the two counts.

Made public

VulnCheck counts all 7,349 ledger entries as "made public" (VulnCheck). Anthropic describes the ledger as hash commitments for findings "we have reported or plan to report". Before a maintainer is notified, an entry shows only the hash and commitment date. The project and bug class appear "only when the disclosure window closes" (CVD dashboard).

CVEs

VulnCheck's 442 CVEs come from its own tracking across CVE records, GHSAs, Linux kernel commits and the ledger (VulnCheck). Anthropic's 733 identifiers count CVE records and GitHub Security Advisories issued for ledger findings, and "a single finding may carry both" (CVD dashboard). VulnCheck writes that it has tracked "more than twice as many" CVEs as the 217 fixed vulnerabilities on Anthropic's ledger (VulnCheck).

Dates and duplicates

VulnCheck's comparison table labels its earlier column "September 8", the date of its previous post. The numbers in that column (26,153 candidates, 91.4%, 2,736 entries, 202 fixed, 245 withdrawn, 2 duplicates) match Anthropic's snapshot dated August 26, 2026 (overview, ledger). In that snapshot the ledger labels the status "merged".

For duplicates on November 2, VulnCheck's text says 1,654 and its table says 1,670 (VulnCheck). Anthropic's ledger shows 1,654 (ledger).

Anthropic's ledger by status, November 2, 2026

These three counts are the ones VulnCheck compares. All three are status filters on the same ledger page, so they are counted the same way.

Anthropic's dashboard over time

Each row is a dated snapshot that Anthropic keeps on its archive page. In each snapshot the true positive rate is findings confirmed valid out of findings reviewed by external firms, as the snapshot shows it.

Anthropic's May 22, 2026 update post gives a separate figure for a narrower group: of 1,752 findings Mythos Preview rated high or critical that were assessed, "90.6% (1,587) have proved to be valid true positives".

Has Anthropic responded?

As of 2026-11-11, we found no public response from Anthropic to VulnCheck's November 9 post on anthropic.com, red.anthropic.com or the CVD dashboard.

The dashboard's About page does explain one change VulnCheck discusses. It predates VulnCheck's post: "The November 2, 2026 update showed three totals lower than in the October 2, 2026 update. They were Total reported to maintainers, Acknowledged by maintainer and Reported direct to maintainers by Anthropic. We found that many reported findings were duplicates of our other findings. We marked them duplicate in the ledger and stopped counting them. We also stopped counting some findings we withdrew."

We will update this page if Anthropic publishes a response.

Related: the AY CLAUDE skills catalog.

Sources

All checked 2026-11-11.

• VulnCheck blog, Patrick Garrity, November 9, 2026: https://www.vulncheck.com/blog/anthropic-glasswing-november-update
• VulnCheck's tracking repo: https://github.com/patrickmgarrity/Anthropic-Credited-CVEs
• Anthropic, Project Glasswing announcement, April 7, 2026: https://www.anthropic.com/glasswing
• Anthropic, "Project Glasswing: An initial update", May 22, 2026: https://www.anthropic.com/news/glasswing-initial-update
• Anthropic, OSS Scanner launch post, October 8, 2026: https://www.anthropic.com/research/launching-opt-in-vuln-finding-service-for-open-source
• Anthropic CVD dashboard, overview, ledger and About page, snapshot 2026-11-02 19:30 UTC: https://red.anthropic.com/2026/cvd/, https://red.anthropic.com/2026/cvd/ledger/, https://red.anthropic.com/2026/cvd/about/
• Anthropic CVD dashboard archive (May 22, August 26 and October 2, 2026 snapshots): https://red.anthropic.com/2026/cvd/archive/

Related guides

• Claude Mythos 5.2 vs Fable 5.2: which to use
• Claude in November 2026: every release and what it means