Anthropic has opened an opt-in vulnerability scanner for maintainers of important open source projects, using its AI models to send security reports more quickly than its human review process can currently manage. In its launch announcement, the company says it found more than 29,000 candidate vulnerabilities during six months of scanning and manually reviewed or triaged about 6,000. The larger figure is a count of possible findings, not 29,000 confirmed, distinct security flaws.
Why Anthropic is offering unreviewed reports
Anthropic says its existing coordinated vulnerability disclosure process uses human review before sending verified reports to maintainers. That creates a bottleneck as automated scanning produces more candidates. Its new OSS Scanner gives projects that choose to enroll an earlier stream of reports generated by models without that initial human triage. The company says it has already sent nearly 5,000 unverified reports in bulk to maintainers who specifically requested everything available. Those reports can include mistakes, duplicate findings, severity estimates that do not fit a project’s threat model, or proposed fixes that need testing.
Each new report is intended to include an explanation, a reproducer, and a candidate patch when available. Anthropic says the service is free for eligible projects and will scan them periodically. It runs scanning agents in hardened sandboxes without internet access after the project environment has been built. A maintainer enrolls by submitting a configuration pull request to Anthropic’s public OSS Scanner repository, including a repository link, contact details, and a Dockerfile for the scan environment. Anthropic says it will verify that the applicant is a core maintainer before accepting a project.
What the numbers do—and do not—show
Anthropic reported a focused check of 97 high- and critical-severity scanner findings across 48 projects. It says external reviewers judged 85 to meet the bar for its coordinated disclosure process; 11 others were real but duplicated known issues or other scan results, and one was invalid. That is encouraging for the selected sample, but it is not an independently established accuracy rate for all 29,000 candidates. The sample was chosen from higher-severity findings, and the company itself warns that raw reports may be wrong.
The launch announcement does not establish that only 516 of the 29,000 candidate findings have been fixed, a figure repeated in some secondary coverage. It also does not provide a denominator of confirmed, disclosed vulnerabilities against which a global fix rate could be calculated. A responsible reading separates candidates, reviewed findings, maintainer reports, and completed patches. Collapsing those categories would turn a real disclosure challenge into a misleading metric about the state of open source security.
What changes for maintainers
For a project with the capacity to triage reports, early access can shorten the time between discovery and a fix. Anthropic cites feedback from maintainers of projects including PostgreSQL, OpenSSL, wolfSSL, and curl, but their experiences should be read as reported examples rather than a promise that every project will see the same quality. Maintainers still need to reproduce the finding, decide whether it crosses a real trust boundary, test the proposed patch, and coordinate disclosure where appropriate.
Anthropic says it will not impose its usual 90-day disclosure period on unvalidated OSS Scanner reports. That clock may begin later if a finding is manually validated under its normal process. Projects can pause reports or opt out. These details matter because a stream of possible vulnerabilities is useful only if the people responsible for fixing them can prioritize it without being forced to treat every model output as a confirmed emergency.
The new service is a test of how AI-assisted discovery fits into the slower work of verification and remediation. Its strongest immediate claim is that it can deliver candidate findings to willing maintainers sooner. The security outcome will depend on how accurately those candidates are triaged, how quickly maintainers can act, and how clearly future reporting distinguishes possible bugs from verified vulnerabilities and fixes.