Anthropic lets Claude hunt for security holes in open source

What's new

Anthropic has launched OSS Scanner, a service that uses Claude to look for security vulnerabilities in open source projects. Open source software is code that is open to everyone, and many companies and public-sector systems build on it. The service is free for participating projects and is governed by special OSS Scanner terms.

It is a sign-up service: projects must enroll themselves. Maintainers do this by opening a pull request (a proposed change) in Anthropic's GitHub repository. There they add a folder containing a configuration file, a Dockerfile and, optionally, a description of the project's threat model. According to ProPakistani, the service is powered by Claude Mythos, which the outlet calls Anthropic's most powerful model. However, the sources do not make clear how the model is formally tied to the service.

According to Cyber Security News, OSS Scanner builds on Project Glasswing. By October 2026, Anthropic had reviewed more than 6,000 vulnerability reports in its existing coordinated disclosure process.

What's smart

The scanner first builds the project in an isolated virtual machine with network access, so that all the components the project uses can be downloaded. The code is then analyzed in an environment with no internet access at all. That is why all dependencies must be included in the Dockerfile.

When something is found, it is sent by email to the project's primary contact. The email includes steps to reproduce the bug and, where possible, a proposed patch, in other words a fix. Reports can be encrypted with a PGP key if the maintainer provides one. Findings are not made public, and no 90-day disclosure deadline applies. According to ProPakistani, projects receive periodic, thorough scans, and Cyber Security News reports that the frequency depends on demand and the project's usage.

There are also validation tools that maintainers can run before submitting their enrollment.

Cheaper or better?

The price is free for participating projects, and no other pricing has been disclosed. The sources do not say whether the service is available separately in Denmark or the EU, and there is no direct comparison with competing tools.

There are important caveats. The reports are AI-generated and not reviewed by humans. According to ProPakistani, Anthropic acknowledges that some findings may be wrong or invalid, so developers must verify them themselves. That matters because a flood of low-quality AI reports burdens maintainers. According to ProPakistani, Linus Torvalds has criticized poor AI-generated security reports, and Google has had problems with AI-generated submissions in its open source vulnerability programs. The same outlet notes, however, that AI-found vulnerabilities such as "Copy Fail" in Linux distributions have shown that AI can be useful in vulnerability hunting.

Access is also limited. According to Cyber Security News, admission requires wide adoption, exposure to remote attacks and importance to infrastructure security, and Anthropic manually verifies applicants' credentials.

What it's good for

The service is aimed at maintainers of widely used open source projects, who are often a few volunteers with limited time to review code thoroughly. For them, a free, recurring scan can be an extra safeguard, especially because it delivers reproduction steps and suggested fixes.

For ordinary users and companies, the indirect benefit is that holes in the software they depend on can be found and patched before attackers exploit them. You don't have to use the service yourself, but you can benefit if the programs you use are enrolled.

A practical tip from Cyber Security News: The email addresses listed in the configuration are public, so projects should use security aliases rather than personal addresses.