Security

Anthropic Launches OSS Scanner: Free AI Vulnerability Scans for Open-Source Projects

Oct 9, 2026·6 min read·By Rahish Saifi

Cover: Anthropic OSS Scanner — free AI vulnerability scans for open-source projects — AI with Rahish

On October 8, 2026, Anthropic announced OSS Scanner: a free, opt-in service that runs periodic security scans on eligible open-source codebases using the company’s strongest models—including Claude Mythos—at no cost to maintainers. The launch is the public face of work informed by Project Glasswing, Anthropic’s effort to apply frontier AI to defensive security, and it deliberately echoes the shape of Google’s OSS-Fuzz: infrastructure-scale help for software the world depends on, not a paid enterprise SKU.

For maintainers of databases, TLS stacks, and other critical libraries, the practical question is how OSS Scanner differs from coordinated vulnerability disclosure (CVD) with human triage—and from Claude Security, Anthropic’s separate enterprise product. OSS Scanner trades human review for speed and frequency: every report is fully model-generated without human review or triage, so teams get findings faster but must treat noise and false positives as part of the deal.

What enrolled projects receive

Projects that pass eligibility and enroll get thorough, periodic scans of their repositories. When the models flag issues, maintainers receive reports that Anthropic says typically include a reproducer, a written explanation (with bisection when possible), and a candidate patch (when available). Early validation covered dozens of projects and produced hundreds of bug reports, including multiple vulnerabilities Anthropic chained into unauthenticated remote code execution exploits, according to Anthropic’s launch post.

SiliconANGLE reports that the Defender Advantage Fund, set up in August, pays to keep the scanner free.

Six months of large-scale scanning

Before opening enrollment, Anthropic says it spent the last six months scanning important open-source software. That work surfaced more than 29,000 candidate vulnerabilities. Researchers manually reviewed roughly 6,000 of those candidates and sent nearly 5,000 reports to maintainers who had asked to receive bulk unverified findings—a different pipeline from OSS Scanner’s ongoing, model-only output, but evidence of how aggressively frontier models can trawl large codebases.

On CyberGym, an academic vulnerability-finding benchmark, LLMs went from finding under 20% of vulnerabilities at the beginning of last year to over 85% this year—context for why the company is productizing scans now.

Expert validation: 88% met the CVD bar

To stress-test quality, Anthropic had an expert penetration tester review 97 critical and high-severity findings across 48 projects. Of those, 85 (88%) met the bar for coordinated disclosure; 11 were real but duplicated known issues or other findings from the scan; and one was an invalid false positive. That ratio is encouraging for an automated pipeline, but it also underscores that maintainers should verify before treating every report as a CVE-ready issue.

How OSS Scanner differs from Claude Security and CVD

Claude Security remains Anthropic’s enterprise-oriented security product. OSS Scanner is explicitly not that: it is a maintainer-facing, no-cost program for open source, with outputs shipped without Anthropic staff validating each finding. That is faster and scales to continuous scanning, but Anthropic warns reports may be incorrect or invalid.

Traditional CVD paths—embargoed review, vendor coordination, CVE assignment—assume human judgment upstream. OSS Scanner inverts part of that: the model does the first pass alone. Anthropic says it will keep sending human-verified reports through its existing CVD process, especially to projects without the resources to triage reports themselves; OSS Scanner is for teams that want a firehose of candidate issues and the tooling to reproduce and fix them locally.

What maintainers are saying

Anthropic quoted several early participants. Noah Misch (PostgreSQL) said several reports “came with fixes we can use nearly as-is”; Anton Arapov (OpenSSL Corporation) said they were “as good and sometimes better than what we get from people”; Eddie Kohler (HotCRP) called them “thorough and clear.” Todd Ouska of wolfSSL reported receiving 74 reports, with all but two valid; five of those became CVEs—useful signal for crypto and TLS maintainers weighing enrollment.

These anecdotes are not a guarantee your project will see the same hit rate; they illustrate that serious libraries are already consuming model-generated reports in production workflows.

How to enroll

Enrollment is opt-in and maintainer-driven. Core maintainers submit a pull request to the anthropics/oss-scanner repository, adding a projects/<name>/project.yaml file. Anthropic’s FAQ and enrollment details are at red.anthropic.com/oss-scanner.

Eligibility is described as similar to OSS-Fuzz: projects should have critical impact on infrastructure and user security, with decisions made case by case. If your library sits on millions of dependency graphs, you are the target audience; small hobby repos may not qualify.

Caveats for maintainers

No human triage: Treat every report as a hypothesis until you reproduce it.
False positives: The expert sample had one invalid finding among 97; your rate may differ by language and codebase age.
Duplicates: Real bugs may already be tracked internally or upstream—budget time to dedupe.
Not a replacement for fuzzing or audit: OSS Scanner complements OSS-Fuzz-style continuous fuzzing and professional review, rather than replacing them.

Related Anthropic cyber programs

The same announcement ties OSS Scanner to wider open-ecosystem support. Claude for OSS offers free Claude Max 20x subscriptions to help maintainers remediate vulnerabilities, while the new Cyber Verification Program gives qualifying security professionals advanced cyber capabilities and reduced blocking classifiers.

The bottom line

OSS Scanner is Anthropic’s bet that frontier models are finally good enough to run continuous, no-cost vulnerability discovery on critical open source—if maintainers accept unreviewed, model-generated reports and do their own validation. It launched October 8, 2026, with enrollment through GitHub and a bar similar to OSS-Fuzz. For infrastructure maintainers, it is a new free layer in the defense stack; for the industry, it is another data point that AI-driven finding on CyberGym-style benchmarks has crossed from experiment to shipped product.