Anthropic Launches Free OSS Scanner After AI Finds 29 000 Vulnerabilities

Anthropic is launching OSS Scanner, a voluntary service in which the company's strongest models, including Claude Mythos, periodically scan open-source projects for vulnerabilities. It is free for projects. The background: over the past six months, under Project Glasswing, the models found more than 29 000 vulnerability candidates in important software projects. The team has managed to manually review and triage about 6 000 of them. The rest is stuck for lack of people who can confirm the findings.

Why a separate service was needed

In a year and a half, the picture has changed radically. On the academic CyberGym benchmark, the share of vulnerabilities found by language models rose from under 20% at the start of last year to over 85% this year. Maintainers who used to get mostly junk from LLMs are now getting quality bugs.

At some point, maintainers started asking Anthropic for everything it had, even unverified findings, along with proposed patches. Almost 5 000 such reports have already been sent directly. The logic is clear: exploits are now written in minutes, and whoever finds and closes a hole faster wins. Attackers are looking for the same weak spots.

Coordinated vulnerability disclosure (CVD) with manual verification is not going away. It remains the main path for projects that lack the resources to triage reports themselves. OSS Scanner adds a fast lane for those who are ready to receive findings immediately.

How the service works

The idea is inspired by Google OSS-Fuzz, which uses fuzzers to find vulnerabilities in open code. OSS Scanner differs from Claude Security, the enterprise product aimed at protecting company systems, in that it offers free audits for open source. The main caveat: results are generated entirely by the model, with no review or triage by humans. That makes scanning faster and more frequent, but it allows for erroneous or invalid reports.

• Models: Anthropic's strongest, including Claude Mythos.
• Participation format: opt-in, periodic scans.
• Report contents: self-contained reproducing code, an explanation of the vulnerability, bisection to the commit that introduced the bug where possible, and a candidate patch.
• Pilot: several weeks across dozens of projects, hundreds of reports, including vulnerability chains that lead to unauthenticated remote code execution.
• Cost to projects: free.

What maintainers say

Feedback from pilot participants confirms the quality. Noah Misch of PostgreSQL notes that an unusually high share of the findings pointed to real defects, and some fixes could be applied almost without changes. The fast track let the team close fresh issues before the GA release. Anton Arapov of OpenSSL Corporation recalls that early AI reports from a year and a half ago were awful. In his assessment, Anthropic's reports, including raw model output, are no worse than human ones and sometimes better, especially when they come with a working exploit that can be checked right away.

wolfSSL's numbers: of 74 reports, all but two were valid, and five became CVEs. Eddie Kohler of HotCRP praised the model's grasp of a complex permissions model and its prioritization of bugs.

For independent validation, Anthropic asked expert pentesters, who normally review findings for CVD, to assess 97 critical and high-severity vulnerabilities from 48 projects. The final figure is cut off in the available text, so it is best to rely on the pilot participants' feedback and the wolfSSL statistics.

What maintainers should do

Projects with an active team and an established process for handling reports should consider opting in. Reports with reproducing code are quick to verify. But decide in advance who will work through the stream and how to filter out false positives, because there is no human review on Anthropic's side. Small projects without the resources for triage are better off staying with classic CVD and manual verification. Everyone else should assume that attackers get the same model capabilities, so the pace of fixes now determines security.