AI Security
Anthropic Launches Free Vulnerability Scanner for Open-Source Projects
The AI company is offering maintainers opt-in access to automated security reports on critical flaws, generated by its strongest models and delivered without human review.
InfoQ
Anthropic has unveiled OSS Scanner, a free vulnerability-detection service designed to help maintainers of high-impact open-source projects identify security flaws in their codebases. The system uses Anthropic's most capable models, including Claude Mythos, to generate periodic security findings and accompanying patch proposals. Maintainers review findings at their own pace and decide which fixes to implement.
How It Works#
The scanner deploys AI agents that independently verify suspected vulnerabilities before sending reports. Each report provides a self-contained test case demonstrating the issue, an analysis of how the flaw was introduced (when possible, through code bisection), and a candidate patch. Projects set up access by submitting a GitHub pull request with build configuration, after which scanning runs offline without requiring internet connectivity.
Maintainers can define a threat model document specifying scope, severity thresholds, untrusted input sources, and reporting preferences. This optional configuration helps the scanner align findings with project-specific security concerns and avoids irrelevant reports.
Scale and Accuracy#
Over six months, Anthropic's models identified over 29,000 candidate vulnerabilities across participating projects. Human review became the bottleneck—researchers manually validated roughly 6,000, yet maintainers who requested early access received nearly 5,000 unreviewed findings with patches.
To assess reliability, Anthropic had penetration testers evaluate 97 critical and high-severity findings across 48 projects. Of these, 85 met established vulnerability disclosure standards. Eleven were real bugs that either duplicated known issues or overlapped with other scanner detections, and one proved invalid. Maintainers have reported occasional overestimation of severity and occasional misalignment with threat models, though early users praised the quality. PostgreSQL's Noah Misch noted that "several reports came with fixes we can use nearly as-is." wolfSSL's Todd Ouska reported that "of the 74 reports we received, all but two were valid, and five became CVEs." Notably, the scanner identified an HTTP/2 use-after-free flaw in curl that could enable remote code execution—since patched in version 8.23.0.
Eligibility and Enrollment#
Anthropic prioritizes established infrastructure projects with the capacity to handle and validate high-severity findings. Prospective maintainers enroll through a GitHub pull request specifying build dependencies; the company manually verifies that applicants are core maintainers and that findings will reach authorized security contacts. Projects can pause or exit the program anytime. Unverified reports carry no automatic 90-day disclosure window, although findings later validated may enter coordinated disclosure.
