On 8 October 2026, Anthropic introduced OSS Scanner, a free, opt-in service for periodic security scans of accepted open-source projects. Unlike its existing coordinated disclosure process, this service sends model-generated reports to participating maintainers without human review. That can shorten delivery time, but it leaves the receiving project to verify and prioritize the findings.
Anthropic says a report can include a reproducer, an explanation, a bisection of when a bug was introduced where possible, and a candidate patch where available. Its service FAQ says automated agents double-check bugs and propose fixes before reports are emailed. Those steps do not amount to human triage, and Anthropic warns that reports can be incorrect.
Enrollment requires maintainer authority and a buildable project
A core maintainer applies through a pull request to Anthropic's OSS Scanner configuration repository. The configuration needs a repository URL, a public primary contact address and a Dockerfile that prepares the project for scanning. Anthropic says it manually verifies maintainer authority and considers eligibility case by case, using criteria similar to OSS-Fuzz for established projects with a critical impact on infrastructure and user security.
The Dockerfile can fetch dependencies during setup, but Anthropic says the audit itself runs without internet access in an isolated environment. Maintainers may provide a threat model and severity preferences to help the scanner interpret project-specific boundaries. The FAQ says projects can pause automated reports or opt out, and recommends the service mainly for teams already able to handle verified high-severity findings.
Faster delivery does not establish validity
In its launch post, Anthropic says outside penetration testers checked 97 selected high- and critical-severity findings across 48 projects, and 85 met its human disclosure bar. That company-reported, selected early sample is not an independently established false-positive rate for future scans. Anthropic also acknowledges duplicate findings, inflated severity ratings and mistakes about a project's threat model.
Anthropic says unvalidated scanner findings carry no automatic 90-day disclosure deadline and will not be made public. If a report is later human-validated through its existing coordinated disclosure process, a disclosure period may begin when the maintainer receives that notice. The service FAQ sets no fixed scan frequency. For maintainers, the immediate decision is whether their triage process can absorb another stream of unverified reports.
