Anthropic has opened OSS Scanner, a free, opt-in service that uses its strongest Claude models, including Claude Mythos, to scan open-source projects for security vulnerabilities. The company announced it on October 8, 2026,. Core maintainers enroll by submitting a pull request, and the service costs them nothing. The main trade-off is that reports go out without human review, so some may be wrong.
What OSS Scanner is
Anthropic’s Frontier Red Team describes OSS Scanner as an opt-in service built on its experience using Claude to find vulnerabilities during Project Glasswing. Enrolled projects receive periodic security scans from the company’s strongest models at no cost. Anthropic says it modeled the service on Google’s OSS-Fuzz, which does similar work with fuzzers rather than language models.
Each report can include a self-contained reproducer, an explanation of the flaw (with a bisection to the change that introduced it, where possible) and a candidate patch when one exists. According to the project’s FAQ, an automated pipeline double-checks bugs, proposes patches and runs root-cause analysis, and a harness checks that a finding concerns the target project rather than one of its dependencies.
How projects get in
Enrollment happens on GitHub. A core maintainer opens a pull request against the anthropics/oss-scanner repository that adds a project configuration file. The FAQ lists three required items: the repository to clone, a primary contact and a Dockerfile that installs every dependency so the agent can work offline. Optional settings include a threat-model file to guide the scan and a disabled: true switch to pause or opt out.
Eligibility is decided case by case, using criteria modeled on OSS-Fuzz. Anthropic says it looks for projects with a critical impact on infrastructure and user security, such as software exposed to remote attacks or one that processes untrusted input. Factors include age, update frequency, the OpenSSF criticality score, an open-source license and usage. Brand-new projects typically do not qualify. Anthropic says a human reviews every sign-up.
Limits and caveats
Anthropic is direct about the downside. Because the output is generated by models and is not reviewed or triaged by people, reports may be incorrect or invalid. Anthropic says some maintainers found severity ratings inflated or said the scanner misread their threat model, while high and critical findings have seldom been called invalid.
There is also no standard 90-day disclosure clock for unvalidated findings. If a report is later validated through Anthropic’s existing coordinated disclosure process, disclosure may begin 90 days after the maintainer is notified.
The numbers Anthropic cites
All figures below come from Anthropic’s launch post and are the company’s own. Over six months its scans produced more than 29,000 candidate vulnerabilities, of which about 6,000 were manually reviewed and triaged, and nearly 5,000 reports went directly to maintainers. In a check of 97 critical and high-severity findings across 48 projects, penetration testers judged 85 (88 percent) to meet the bar for coordinated disclosure; of the other 12, 11 were real but duplicated other findings and one was a false positive. We have not independently verified these figures.
Why it matters
The announcement lands in a month dominated by AI-found bugs. PHP2Shell showed frontier models from other labs uncovering decade-old flaws that were quickly exploited, and Anthropic itself has warned about small models and IoT worms. OSS Scanner is the defensive counterpart: an attempt to get findings to maintainers before attackers reach them. Our explainer on AI and cybersecurity covers both sides.
For maintainers, the practical question is capacity. Unreviewed reports are only useful if a project can triage them, and Anthropic says the service is meant for projects already able to handle verified high and critical reports.
What to watch next
- How many projects are accepted and whether the eligibility criteria change as Anthropic says they may.
- Whether maintainers report that false positives or inflated severities ease as models improve.
- Whether other labs offer comparable opt-in scanning for open-source projects.
Key takeaways
- OSS Scanner is a free, opt-in service from Anthropic that scans enrolled open-source projects with its strongest models, announced October 8, 2026.
- Core maintainers enroll with a pull request to the anthropics/oss-scanner repository; Anthropic reviews each sign-up.
- Reports go out without human review, so some may be incorrect or have inflated severity.
- Anthropic cites 29,000 candidate vulnerabilities over six months and an 88 percent validation rate on a 97-finding sample; these are the company’s own numbers.
- Eligibility follows OSS-Fuzz-style criteria and is decided case by case.
Frequently asked questions
Is Anthropic’s OSS Scanner free?
Yes. Anthropic says enrolled projects receive scans at no cost and that it covers the full cost itself.
How do I enroll a project?
A core maintainer opens a pull request on the anthropics/oss-scanner GitHub repository with a project configuration file. It must name the repository, a primary contact and a Dockerfile that builds the project.
Are the vulnerability reports checked by people?
Not before delivery. Anthropic says reports are model-generated without human review, so some can be wrong; only findings later validated go through its coordinated disclosure process.
