Anthropic's Free OSS Scanner Sends Unreviewed Reports

What Launched On 8 October

Anthropic launched the Anthropic Cyber Mission on 8 October with two programmes: OSS Scanner for open-source projects and the Critical Infrastructure Defense Program for the providers that secure operational technology.

The infrastructure programme brings Claude models, on-site engineers and threat research to security providers that protect power grids, water systems and transport networks. Its founding partners are Accenture, Booz Allen, CrowdStrike, Deloitte, Dragos, Hitachi, Insane Cyber, Nozomi Networks, Palo Alto Networks, PwC and Rockwell Automation. Anthropic says it will start with a small cohort of providers.

OSS Scanner is an opt-in service inspired by Google's OSS-Fuzz. Enrolled projects receive periodic security scans from Anthropic's strongest models at no cost, and The Verge reports those models include Claude Mythos. Each report carries a proof of concept, an explanation and a suggested fix where one is available.

Reports Sent Without A Human In The Loop

OSS Scanner's reports are model-generated and sent to maintainers without human review or triage, which Anthropic says makes them arrive faster and means some will be wrong.

Anthropic's own example of an error is an incorrect severity rating. It expects a true-positive rate above 90 percent, which leaves fewer than one report in ten wrong on its own estimate, and a maintainer cannot tell in advance which one.

Enrolment is by pull request from a project's core maintainers, using criteria close to Google's OSS-Fuzz: critical impact on infrastructure and user security, exposure to remote attacks and the number of projects that depend on it. Anthropic says plainly that many projects are already overwhelmed by the reports they get and that the service is built for projects that can already keep up with verified high and critical vulnerability reports.

Unvalidated findings carry no 90-day disclosure clock, because Anthropic does not want to force maintainers to read findings carefully when it has not done so itself. A project can pause reports with one configuration line or leave, and it then returns to the standard path, where Anthropic says it had reviewed over 6,000 reports by October 2026.

Why Fixing Is Now The Bottleneck

Anthropic's own launch post says it is easier than ever to find vulnerabilities, while verifying, prioritising and fixing them remains hard.

In Project Glasswing, Anthropic says, months often passed between a vulnerability being found and being fixed. For operational technology a fix may have to wait until it can be applied safely to running machinery, in rare cases for decades. Anthropic forecasts that in two years AI will favour defence, but says that in the near term it may not, because the cost of exploiting vulnerabilities has dropped while verifying and fixing still depend on people.

The Verge adds the pressure at the receiving end: some open-source projects, including Linus Torvalds and Google, are struggling to keep up with the sudden onslaught of AI-generated bug reports. A free scanner adds reports of its own for every project that enrols, which is why Anthropic limits it to projects with the capacity to keep up.

Anthropic says it has funded the organisations behind widely used open-source code, among them the Python Software Foundation and the Apache Software Foundation, plus Alpha-Omega and OpenSSF through the Linux Foundation. It adds that a Defender Advantage Fund launched in August keeps OSS Scanner free.

What To Do With Your Dependency List

A company that depends on open-source libraries should not read this launch as a safer supply chain, because the scanner reaches only projects that opt in and have the capacity to triage.

Ask the open-source suppliers you rely on most whether they enrolled, who will read the reports and what response time you can expect. If you maintain open source yourself, decide before enrolling who triages the reports and what happens to the wrong ones. If you only consume it, budget patch time or fund the maintainers, because faster discovery raises the number of fixes you will have to apply.

The EU Cyber Resilience Act supplies the clock on the manufacturer side. The European Commission says manufacturers must report actively exploited vulnerabilities from 11 September 2026, with an early warning within 24 hours of becoming aware, a full notification within 72 hours and a final report no later than 14 days after a fix is available. Open-source software stewards face reporting duties from 11 December 2027 and, under the Act's light-touch regime for them, are not subject to administrative fines.

If you run operational technology, ask your security provider whether it is a founding partner of the programme, since the Critical Infrastructure Defense Program brings Claude models, on-site engineers and threat research to those providers, and Anthropic says critical infrastructure is hard to defend in many ways that AI cannot fix.

Read next: An AI Agent's False Murder Tip Sat Unseen for 72 Days  |  Google Freezes Open-Source Bug Bounty Intake