Overview
OSS Scanner is the opt-in vulnerability scanning service Anthropic launched on October 8, 2026 for open-source projects, running periodic scans with its strongest models including Claude Mythos at no cost. The idea comes straight from Google OSS-Fuzz: continuous automated scanning of open source proved its worth with fuzzers, and Anthropic wants to run the same play with language models.
The backdrop is that models changed character over two years. Anthropic cites CyberGym, an academic vulnerability-finding benchmark, where the share of vulnerabilities models can find went from under 20 percent early last year to over 85 percent this year. What maintainers receive changed with it, moving from mostly slop to high-quality bug reports. Over the past six months Anthropic used its latest models to scan some of the most important software projects in the world and discovered over 29,000 candidate vulnerabilities, but human capacity to validate them only covered about 6,000. That bottleneck is the direct reason OSS Scanner exists: rather than letting reports queue for human triage, open a fast track for projects willing to consume raw findings themselves.
To be clear about what that fast track means, its output is fully model generated with no human review or triage. What that buys is faster and more frequent scanning, and the trade-off is that reports may be incorrect or invalid. Anthropic keeps its existing human-verified coordinated vulnerability disclosure process running in parallel, so both paths stay available.
Key Features
- Free, periodic, built for open source: Periodic security scans run by Anthropic strongest models including Claude Mythos at no cost to open-source projects. It sits on a separate track from Claude Security, which helps enterprises defend their own systems, while this one audits open-source projects
- Every report ships with a reproducer and a patch: Each report contains a self-contained reproducer, an explanation of the vulnerability, a bisection identifying when the bug was introduced where possible, and a candidate patch when one is available, so verification can start immediately instead of reconstructing the scene first
- Opt-in enrollment: Core maintainers of eligible projects enroll by submitting a pull request to a designated GitHub repository following a standard project template, with extended guidance in a FAQ. Eligibility follows criteria similar to OSS-Fuzz, requiring critical impact on infrastructure and user security, decided case by case
- Accuracy numbers from early validation: To validate an early version, Anthropic had the expert penetration testers who review its CVD findings check 97 critical and high-severity findings across 48 projects. Of these, 85, or 88 percent, met the bar for disclosure, and of the remaining 12, eleven were real but duplicated known issues, leaving a single invalid finding
- Built on Project Glasswing experience: The service grows out of internal practice using Claude to find vulnerabilities. The 29,000 candidate vulnerabilities accumulated over six months, and maintainers repeatedly asking for bulk delivery of raw reports, shaped what the product looks like
- Works alongside companion programs: The Cyber Verification Program makes advanced cyber capabilities and reduced blocking classifiers available to qualifying security professionals, while Claude for OSS provides free Claude Max 20x subscriptions to help remediate vulnerabilities and improve open-source projects
Use Cases
- [object Object]
- [object Object]
- [object Object]
- [object Object]
- [object Object]
Pros
- Completely free for open-source projects, scanned by Anthropic strongest models
- Every report carries a self-contained reproducer and candidate patch, keeping verification cheap
- 88 percent of critical and high-severity findings met the disclosure bar in early validation, with a single false positive
- Enrollment needs only one pull request, following mature eligibility criteria modelled on OSS-Fuzz
- Runs alongside the human-verified CVD process, so projects choose between fast track and human triage
- Claude for OSS supplies free Claude Max 20x subscriptions, so resources exist for the remediation phase too
- Scanning is periodic rather than one-off, keeping pace with code evolution
Pricing
Free for eligible open-source projects with no scanning fee. Projects enroll when a core maintainer submits a pull request, and eligibility is assessed case by case on critical impact to infrastructure and user security. Reports are model generated without human review, so projects verify and triage them themselves. Claude for OSS complements the service with free Claude Max 20x subscriptions for open-source work.
Summary
What OSS Scanner actually fixes is a throughput mismatch: models find vulnerabilities far faster than humans can verify them, and Anthropic is sitting on 29,000 candidates with only 6,000 triaged. Rather than letting the remainder queue indefinitely, hand them to maintainers who are willing to do their own triage.
For open-source projects that is a good deal, free, periodic, with reproducers and patches attached, and an 88 percent bar-meeting rate on critical and high-severity findings showing the signal is solid. What matters to remember is its nature: fast-track output is raw model output with no human backstop, so reports still need your own verification, and severity ratings may run high or misread a project threat model. Anthropic kept the human-verified CVD path precisely for projects without the resourcing to triage themselves.
If your project is widely depended-upon infrastructure, it is worth having a core maintainer open that pull request.
Version History
- OSS Scanner opens for enrollment (2026-10-08): Anthropic launched this opt-in vulnerability scanner for open-source projects, inspired by Google OSS-Fuzz and powered by its strongest models including Claude Mythos at no cost. Output is fully model generated without human review or triage, with each report containing a self-contained reproducer, an explanation, a bisection to the introducing change and a candidate patch. Core maintainers enroll via pull request, with eligibility assessed case by case on OSS-Fuzz-like criteria. Early validation found 85 of 97 critical and high-severity findings across 48 projects met the CVD disclosure bar, with one invalid result. The launch also introduced the Cyber Verification Program and Claude for OSS, which offers free Claude Max 20x subscriptions to open-source projects