The report
Anthropic announced the opt-in OSS Scanner on October 8, 2026. Eligible open-source maintainers can request periodic model-generated security reports without paying for the scans. Each report may contain a reproduction, an explanation, and a suggested patch, but the new fast-track sends them without individual human review. That is a different workflow from Anthropic's coordinated disclosure of human-triaged findings.
Three counts, three different denominators
Anthropic says its models produced more than 29,000 candidate vulnerabilities over six months and that its staff manually reviewed roughly 6,000. It separately reports having sent nearly 5,000 unverified reports to maintainers who requested bulk delivery. None of those counts should be relabeled as independently confirmed exploitable defects. For a smaller, selected high-severity validation, its external-to-the-model security reviewers examined 97 high/critical candidates across 48 projects: 85 met the team's coordinated-disclosure bar, 11 were real but known or duplicate issues, and one was invalid. These numbers describe that selected validation exercise, not the precision or false-positive rate of the whole 29,000-candidate population.
The receiving end of the pipeline
The launch report offers a useful maintainer-side observation: wolfSSL's Todd Ouska says that 72 of 74 reports received were valid and five became CVEs. This is an attributed testimonial about one project, not a representative cross-project outcome. Anthropic also says some maintainers considered severity ratings inflated or inconsistent with their threat models. A candidate patch therefore still needs code review, reachability testing, regression checks and a maintainer's own impact assessment.
The model output may shorten the time to a testable vulnerability hypothesis, especially if it supplies a reproducer, but the workload does not vanish. Maintainers must decide whether an input is reachable in supported configurations, whether its severity applies to their project, whether a generated patch preserves behavior, and how to handle coordinated disclosure. The service changes who sees candidate reports and when; the announcement does not establish successful attacks on the participating projects or scanner-wide accuracy.
Anthropic says only core maintainers of eligible projects can enroll using a pull request to its enrollment repository. The scope is therefore not an unrestricted public scanning endpoint. Source: https://www.anthropic.com/research/launching-opt-in-vuln-finding-service-for-open-source