Anthropic OSS Scanner audits open source code for free

Anthropic will scan open source projects for security flaws at no charge. Projects that opt in get recurring passes from the company's strongest models, and nothing in the loop waits on a human reviewer. The service is called OSS Scanner.

What lands in a maintainer's inbox

Every report bundles three things: a self-contained reproducer, a candidate patch when the model can produce one, and a bisection pinning the commit that introduced the flaw. The framing matters more than the scanner itself. Anthropic's argument is that discovery stopped being the bottleneck a while ago, and that verifying, triaging and repairing is where the labor still sits.

Skipping human pre-review is a deliberate trade. Maintainers hear about a flaw sooner, and they also inherit the model's mistakes — the company specifically flags severity ratings that come back wrong. A report is a lead, not a verdict.

The hit rate so far

The penetration testers who vet Anthropic's coordinated disclosures went through 97 critical and high-severity findings from an early version, spread across 48 projects. They cleared 85 for disclosure. Of the remaining 12, all but one were genuine bugs that duplicated either known issues or other findings from the same scan.

One maintainer has published its own tally. wolfSSL, which builds an encryption library for embedded systems, received 74 reports during early trials; all but two held up, and five became CVEs.

Getting in

Enrollment runs through GitHub, with core maintainers opening a pull request against an Anthropic repository. Eligibility borrows the OSS-Fuzz bar — projects of "critical impact on infrastructure and user security" — and each case is judged on its own.

The bill goes to the Defender Advantage Fund, set up in August 2026, which has also backed the Python Software Foundation, the Apache Software Foundation, Alpha-Omega and the OpenSSF.

Grids and water systems get a separate program

The same day, Anthropic opened a critical infrastructure effort that puts frontier models and onsite engineers inside power and water utilities. 11 providers signed on as founding partners, spanning consultancies, operational-technology security vendors and industrial equipment makers. The target is the gear in plants and substations that runs for decades and cannot simply be taken offline to apply a patch.

Andrew Turner, who leads commercial cyber at Booz Allen, calls that operational technology "the next frontier for autonomous AI-enabled attacks." Several partners already lean on Claude to close vulnerabilities. The predecessor, Project Glasswing, ran from April to October 2026 and gave vetted organizations access to Claude Mythos; it now folds into an expanded Cyber Verification Program.

What nobody has priced

The infrastructure program's commercial terms are blank. Per the report, Anthropic has not said whether access is free or who pays for the compute. The company expects AI to tilt toward defenders within two years — even as the cost of mounting an attack keeps dropping and confirming a fix stays slow.