Skip to content

Anthropic

Anthropic’s OSS Scanner skips human review on bug reports

5 min read

Written by an AI model from the linked sources and checked against them by a second model. How we work

Image: anthropic.com

Anthropic launched the Anthropic Cyber Mission on October 8 with two programs. One is OSS Scanner, which gives opt-in open-source projects free, periodic security scans from Anthropic’s most capable models. The other is the Critical Infrastructure Defense Program, which works with the security firms that look after power, water and transport systems. OSS Scanner reports reach maintainers with no human review, and Anthropic expects a true-positive rate above 90%. Anthropic says the service is meant for projects that can keep up with what it finds, and that projects without that capacity will keep getting human-verified disclosures under its coordinated disclosure policy.

What OSS Scanner sends

The idea came from maintainers. Under Project Glasswing, Anthropic scanned hundreds of widely used open-source projects. Humans triaged many of the candidate bugs, and the findings were reported privately to the maintainers. Some maintainers who could triage at scale then asked for everything the models had found in their project, reviewed or not. OSS Scanner is the result. Anthropic says it is modeled on Google’s OSS-Fuzz.

Each report has three parts: a proof of concept showing how the bug could be exploited, an explanation, and a suggested fix where one exists. Skipping review gets reports out faster, and Anthropic says some will contain errors, a wrong severity rating among them. A true-positive rate just above 90% still means as many as one finding in ten may not be a real bug. Anthropic says it will work on that rate and on fix quality over time. Core maintainers of critical open-source projects can enroll, and the scanner team has published a separate post on how it works and how to sign up.

The capacity requirement is a direct response to what happened with Glasswing’s disclosures. By May 2026, Anthropic had disclosed about 1,596 vulnerabilities to open-source maintainers, and only 97 had been patched, according to the Cloud Security Alliance. The CSA calls this a structural mismatch between how fast AI finds flaws and how fast maintainers can fix them. Given those numbers, sending unreviewed reports only to projects that ask for them, and only to those that can absorb them, is the sensible way to run the service.

Anthropic’s stated next steps for open source are wider rollout, automated triage and patching for projects that want it (both are still mostly manual), and research into hardening or rewriting code. The Defender Advantage Fund (0xDAF), launched in August, pays for the pilots and keeps the 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 that maintainers aren’t flooded. Maintainers can separately apply to Claude for Open Source for free Claude Max subscriptions.

Critical infrastructure goes through vendors

The infrastructure program doesn’t sell to utilities directly. It brings frontier Claude models, Anthropic engineers working on-site and Anthropic’s threat research to the small group of providers that operators already trust. The founding partners are Accenture, Booz Allen, CrowdStrike, Deloitte, Dragos, Hitachi, Insane Cyber, Nozomi Networks, Palo Alto Networks, PwC and Rockwell Automation. The list includes consultancies, security vendors that protect business and industrial networks, and manufacturers that build and patch the equipment.

Anthropic says several partners already use Claude to fix vulnerabilities and to help their customers do the same. It calls this first group a small cohort meant to show which approaches are effective and practical. Security vendors, system integrators and equipment manufacturers that build products for critical infrastructure can register interest, and Anthropic plans to add more partners and sectors over the coming months. The program builds on the cyber defense program Anthropic launched in June for state, local, tribal and territorial governments. Since then, Anthropic says, it has given Claude models and technical support to more than half of US states and to some of the largest public infrastructure operators in the country.

Anyone else on the defensive side, whether a utility security team or a solo maintainer, can apply to the expanded Cyber Verification Program for access to the models. Earlier the same week, Anthropic folded Project Glasswing into that program.

Finding bugs is no longer the hard part

Anthropic is direct about why it is doing this. Glasswing partners found many vulnerabilities, but the company says cyber risk has not come down enough, because verifying, prioritizing and fixing the findings is still slow. In Glasswing, months often went by between finding a vulnerability and fixing it. Glasswing started on April 7 around Claude Mythos Preview, an unreleased model built for autonomous security research. CNBC reported Anthropic saying that partners had found more than 10,000 high- or critical-severity flaws since the launch.

Attackers don’t have to wait. A Mythos-found authentication bypass in Rejetto HFS was exploited in the wild a day after it was disclosed in early October. Anthropic’s own forecast is that within two years AI will favor defenders, making it easier to catch bugs before they ship and to write secure software from the start. It adds that “in the near term, that may not be true,” because exploiting bugs has become cheap while disclosing and fixing them still depends on people.

The gap is widest for industrial systems. Controllers in plants and substations often can’t be taken offline for patching, so known vulnerabilities can stay open for years. “With operational technology, a fix may have to wait until it can be applied safely to running machinery—in some rare cases, this might take decades,” Anthropic writes.

No comments yet. Be the first.

Join the conversation

Sign in to leave a comment. It appears after moderation.