Anthropic’s free OSS Scanner skips human review on purpose

Anthropic on October 8 opened OSS Scanner, a free, opt-in service that runs its most capable models over enrolled open-source projects and sends the findings to maintainers with no human review. It is one half of the new Anthropic Cyber Mission, the company’s effort to secure open-source software and critical infrastructure. Enrolled projects get reports faster but unchecked: Anthropic expects a true-positive rate above 90% and warns that some reports will contain errors, such as a wrong severity rating. Projects without the capacity to triage at that pace keep getting human-verified disclosures.
What lands in the inbox
Enrolled projects get periodic scans. Each report contains a proof of concept showing how the bug could be exploited, an explanation, and a suggested fix where the model can produce one. The service is modelled on Google’s OSS-Fuzz, and the team’s post explains how it works and how to enroll.
The idea came from maintainers. Under Project Glasswing, Anthropic scanned hundreds of widely used open-source projects, had humans triage and review many of the potential vulnerabilities, and reported what it found privately through its coordinated vulnerability disclosure (CVD) process. Some maintainers who could triage at scale asked for everything the models had found in their project, reviewed or not. OSS Scanner is that request turned into a service.
The 90% figure is still Anthropic’s expectation, not a measured rate, and the company says it will work to improve both the true-positive rate and the quality of the suggested fixes over time.
Who gets in
Anthropic assesses projects case by case, with criteria similar to OSS-Fuzz’s, and gives priority to projects with critical-infrastructure or security implications. Remote exposure and the number of dependent users or projects are among the considerations. The service is aimed at projects that already handle verified high and critical reports and can check additional findings on top of them. The human-reviewed CVD route stays open for projects that need or prefer verified findings.
Anthropic’s own explanation for the change is that Glasswing turned up many vulnerabilities but “we haven’t yet achieved a sufficient reduction in cyber risk.” Finding bugs is easier than ever, the company says, but verifying, prioritising and fixing them is still hard. The gap has already shown up in practice. VulnCheck counted 286 CVEs from Mythos and Glasswing in early October, and attackers targeted one of them, a Rejetto HTTP File Server flaw, a day after a researcher posted a video exploiting it, per The Register.
Anthropic lists three next steps for its open-source work: bring OSS Scanner to more projects while keeping human-verified disclosure for those that need it, automate triage and patching (both still mostly manual) for projects that want that, and research ways to harden or rewrite code for projects that want to go beyond patching. It says it will follow guidance from maintainers and foundations on all three.
Funding and access
The Defender Advantage Fund (0xDAF), launched in August, pays for pilot programs in these areas and keeps OSS Scanner free. Anthropic has also funded the Python Software Foundation, Alpha-Omega and OpenSSF through the Linux Foundation, and the Apache Software Foundation. It supports Akrites and Gold Eagle too, which collect and coordinate vulnerability reports from many sources so maintainers don’t get buried. Maintainers can apply to Claude for Open Source for free Claude Max subscriptions, and to the Cyber Verification Program for wider access to Claude’s cyber capabilities for defensive work. Earlier the same week, Anthropic merged Project Glasswing into that expanded program.
The infrastructure half
The other launch, the Critical Infrastructure Defense Program (CIDP), covers operational technology: the controllers, control software and industrial networks that run power grids, water utilities, factories and transport. These systems often can’t be taken offline for patching, so known vulnerabilities can stay open for years, and Anthropic itself notes that “a mistake can take down a plant.”
Anthropic isn’t selling CIDP to utilities directly. The program brings frontier Claude models, on-site Anthropic engineers and threat research to the providers operators already rely on. The founding partners are Accenture, Booz Allen, CrowdStrike, Deloitte, Dragos, Hitachi, Insane Cyber, Nozomi Networks, Palo Alto Networks, PwC and Rockwell Automation. Several are already using Claude to fix vulnerabilities and helping their customers do the same. Anthropic is starting with a small group of providers to learn which approaches work. A separate program for US state, local, tribal and territorial governments, launched in June, has since offered models and technical support to more than half of all US states.
Companies that build security products or services for critical infrastructure can register their interest through the announcement. Operators reach the program through those providers.
Up next
OpenAI’s 722 math papers come with no one to write to
OpenAI Support declined to forward a mathematician’s complaint about Families 335 and 336, which the mathematician says build on their arXiv paper without citing it. Support couldn’t share any correspondence address either.
Related stories
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

