OSS Scanner: Anthropic’s service for finding vulnerabilities in open source

OSS Scanner: Anthropic’s service for finding vulnerabilities in open source

On October 8, Anthropic launched OSS Scanner, a service that checks open-source projects for vulnerabilities using the company’s strongest models. According to Anthropic, the service builds on its experience using Claude to find vulnerabilities during Project Glasswing.

Anthropic already scanned open-source software before and sent reports through its coordinated vulnerability disclosure (CVD) process, but only after human review. By October 2026, it had reviewed more than 6,000 such reports. Manual review takes time, so findings reached maintainers with a delay. OSS Scanner skips this step: enrolled projects receive reports as soon as a scan is complete.

The service is free to join. To enroll, a project submits a pull request to the anthropics/oss-scanner repository that adds a project.yaml file with the following fields: • repo — a link to the project repository; • primary_contact — the email address that receives reports; • dockerfile — the path to a Dockerfile that builds the environment with all required dependencies. The build runs with network access, while the agents analyze the code in isolated sandboxes with no internet access. Maintainers can also attach a threat model specifying what to test, which inputs to treat as adversarial and how to classify the severity of findings.

The pipeline includes agents that double-check bugs, propose patches and perform root cause analysis. It also verifies that each vulnerability is in the project itself rather than in its dependencies. After the initial scan, projects are rescanned periodically for newly introduced and previously missed vulnerabilities.

Because unvalidated reports may contain false positives, Anthropic does not apply a 90-day disclosure deadline to them and does not publish them. If a human later validates a finding through the standard CVD process, the 90 days start from the moment the maintainer is notified of the validation.

Not every repository is eligible. The criteria are similar to those of OSS-Fuzz: the project’s age and activity, its OpenSSF criticality score, an open-source license and how widely it is used. New projects and most forks usually do not qualify. Anthropic also acknowledges that many projects are already overwhelmed by the volume of reports, so the service is aimed at those who can keep up with verified high and critical vulnerabilities.

What sets it apart from the commercial Claude Security product is that, for open source, Anthropic additionally uses experimental and more token-hungry harnesses to find deeper bugs.