Anthropic publishes a dashboard that tracks how vulnerabilities in open source software (OSS) found by its Claude AI models are being reported to maintainers and fixed. According to the figures as of November 2, 2026, the company has reported a total of 5,087 vulnerabilities to 680 projects, and to its knowledge 629 of them have been patched. The candidates Claude surfaced number 37,966, and the dashboard also makes clear that human review has not kept up.

Of 37,966 Candidates, 6,376 Were Reviewed by Outside Firms

The dashboard covers vulnerabilities found between November 1, 2025 and November 2, 2026, and shows counts at each stage from discovery to fix. The first stage, "candidates," counts every crash or vulnerability hypothesis produced by Claude models before any triage. The models include Claude Mythos Preview as well as other Mythos-class models and other Claude models.

StageCount
Candidates (before triage)37,966
Reviewed by external security firms6,376
Confirmed as real vulnerabilities6,009 (94.2%)
Reported to maintainers (total)5,087
Acknowledged by maintainers3,635
Patched upstream629
CVEs / GHSAs assigned733

The external review was carried out by six firms: Ada Logics, Anvil, Calif.io, Doyensec, Ophion Security and Trail of Bits. Of the 6,376 findings they reviewed, 6,009 were confirmed as real vulnerabilities, a true-positive rate of 94.2%. A total of 733 identifiers have been issued: 230 CVE records and 503 GitHub Security Advisories (GHSAs), and a single finding may carry both.

More Than 60% of Reports Went Out Without External Review

The 5,087 reports to maintainers came through two paths. 1,862 passed through review by an outside firm, while 3,225 were sent directly by Anthropic. The company explicitly notes that the latter may contain false positives. It says direct disclosure is used, for example, when maintainers specifically ask to receive findings that have not been triaged.

The dashboard also comes with caveats on how to read the numbers. The true-positive rate includes real bugs that had already been reported through other channels, as well as findings the maintainers decided not to fix. For that reason, the company describes the true-positive count as only one proxy for impact, and patches as a more reliable one, though only a lagging indicator.

In the November 2 update, three totals, including the number of reports to maintainers, came in lower than in the October 2 update. Anthropic explains that this is because duplicate reports and withdrawn reports were removed from the count.

Human Review Sets the Pace

On why the number of reports is small compared with the number of candidates, Anthropic says independent human triage and review has become the rate-limiting step. Many confirmed vulnerabilities have not yet been reported because of capacity limitations, according to the company. It also points to maintainers receiving a higher volume of inbound security findings.

Unfixed vulnerabilities are recorded without revealing their details. The ledger lists 7,349 entries, each recorded as a SHA-3-512 hash, a fixed-length value computed from the contents. Before maintainers are notified, only the hash and the date it was logged are made public, so the contents can be checked against it later.

Part of Project Glasswing, Launched in April

The effort is part of Project Glasswing, which Anthropic launched on April 7, 2026. The project uses Claude Mythos Preview, a model that is not generally available, for defensive security work, and its launch partners included Amazon Web Services, Apple, Google, Microsoft, NVIDIA and the Linux Foundation. At launch, the company committed up to $100 million in usage credits for Mythos Preview and a total of $4 million in donations to open source security organizations.

What Changes for OSS Maintainers and Users

For OSS maintainers, Anthropic launched OSS Scanner on October 8. Enrolled projects are scanned periodically and at no cost by the company's strongest models, with the results delivered by email. However, reports sent through this path have not undergone human review. The service is aimed at projects that can already keep up with verified high and critical severity reports, and forks and brand-new projects typically do not qualify.

Meanwhile, for those running products and systems built on OSS, the key point is what "patched" means on the dashboard. According to the company, it means a patch landed upstream after the report was sent, and does not guarantee that the patch has been widely installed.

Our Take

The numbers show that AI-driven vulnerability hunting has reached a stage where its volume exceeds human capacity to process it. Against 37,966 candidates, 6,376 have completed external review and 629 have confirmed patches, less than 2% of the candidates. The 94.2% true-positive rate is a credit to the accuracy of the findings.

That said, direct reports without outside review make up more than 60% of the total, and the burden on the OSS maintainers receiving them is likely not small. Still, publishing a hash-based ledger and even corrections to its own totals is an attempt to show progress in a verifiable way. Companies and developers that rely on OSS will likely need to pay closer attention than before to new fixed releases of the projects they use, and to applying them.

Summary

Anthropic disclosed that it has reported 5,087 Claude-found vulnerabilities to 680 OSS projects and confirmed patches for 629 of them. The development matters to OSS maintainers and to companies using products that incorporate OSS. The dashboard keeps earlier totals available by date, so future changes in the numbers can be tracked as well.