Anthropic OSS Scanner Sends Free Bug Reports to Maintainers

Free scans, with verification handed to maintainers

Anthropic launched OSS Scanner on 8 October, offering eligible open source projects free, periodic vulnerability scans by its strongest Claude models. The service sends findings straight to maintainers without human review. That choice can shorten the wait for a useful report, but it also moves the work of checking whether a bug is real, how serious it is, and whether the proposed patch works onto the receiving project. Anthropic's enrollment guide says the service is intended for maintainers who can already handle verified high and critical reports and want to receive more findings.

The launch followed Anthropic's Project Glasswing scanning effort. Anthropic says its models found more than 29,000 candidate vulnerabilities in six months, while staff manually reviewed about 6,000. Those are the company's counts of candidates and reviewed reports, not 29,000 confirmed security flaws. The new opt-in channel is its attempt to deliver some findings before the review queue catches up. Projects that do not enroll can still receive findings through Anthropic's existing human-verified disclosure process.

Early tests offer some evidence of usefulness, with limits. Anthropic says outside penetration testers examined 97 scanner findings marked high or critical across 48 projects. Eighty-five met its disclosure bar, 11 were real but duplicated other findings or known issues, and one was invalid. That sample was selected from high-severity candidates and assessed within Anthropic's programme; it does not establish the accuracy of every future report or the workload for every maintainer. CyberScoop corroborated the programme's launch and its unreviewed reports; the accuracy sample still comes from Anthropic.

Who can enroll and what arrives

Core maintainers apply through a pull request to Anthropic's public oss-scanner repository. They provide a project.yaml with the repository address and a contact email, plus a Dockerfile that installs dependencies and builds the software for analysis. Anthropic says it will manually check that an applicant is a core maintainer. It uses criteria similar to OSS-Fuzz, favouring established projects that affect infrastructure or user security, especially those exposed to untrusted input or used by many other projects. Acceptance remains case by case.

The Docker build can fetch dependencies, but Anthropic says the scanning agents run without internet access afterward. Maintainers can add a threat model to identify intended security boundaries, out-of-scope behaviour, and their severity rubric. That is a practical step, because a model that misunderstands a project's deployment or trust assumptions can write a plausible but misleading report. The repository's contact email is public. Projects can supply a public PGP key to encrypt emailed reports, although that option cannot be combined with extra copied recipients.

After the first scan, Anthropic plans recurring scans and emails bundles of reports. Its research post says a report can include a reproducer, an explanation of the vulnerability, the commit where it arose when that can be determined, and a candidate patch when available. The FAQ does not promise a fixed scan interval. A maintainer should treat the reproducer as a test case, verify it in the project's supported environment, then inspect the patch for regressions before merging it. A suggested severity is a starting point for the project's own assessment.

No automatic public disclosure deadline

Anthropic says unvalidated OSS Scanner reports do not start its usual 90-day public disclosure clock and will not be published by the scanner. If its staff later validate a finding through the separate coordinated disclosure programme, that process can start a 90-day period from the human-validated notification. Maintainers can pause automated reports or remove their project through a repository change.

This arrangement makes the trade-off unusually explicit. A team with security triage capacity may prefer earlier access to candidate bugs, including reports that turn out to be duplicates or have inflated severity. A small project already struggling with its inbox may be better served by the slower reviewed channel. Neither route replaces the project's own testing and release process. OSS Scanner is a new source of leads, with its value depending on whether the reports produce fixes faster than they consume maintainer time.