Anthropic's OSS Scanner: What Free AI Vulnerability Scanning Actually Changes for Open Source

Detail	Finding
Launch date	October 8, 2026, as part of Anthropic's new Cyber Mission
What it is	OSS Scanner, a free, opt in service that scans enrolled open source projects for vulnerabilities
Model	Anthropic's most capable models, including Claude Mythos
Review process	Fully model generated reports, no human review before a maintainer receives them
Expected accuracy	Anthropic states a true positive rate above 90 percent
Scale so far	116 pull requests submitted for enrollment, more than 29,000 candidate vulnerabilities identified, over 6,000 flaws reported to maintainers, 584 advisories published as of October 2, 2026
Inspiration	Modeled on Google's OSS Fuzz, but using model reasoning instead of fuzzing

Anthropic's Cyber Mission announcement covers a lot of ground: a Critical Infrastructure Defense Program with eleven founding partners including Accenture, CrowdStrike, Palo Alto Networks, and Dragos, plus continued outreach offering Claude and technical support to more than half of US state governments since June. But the piece engineering teams who touch open source code should actually read closely is OSS Scanner, because it is live today, it is free, and it already has real numbers behind it rather than a roadmap.

What OSS Scanner actually does

OSS Scanner runs periodic, automated security audits against enrolled open source repositories using Anthropic's strongest available models. Each scan produces a report that includes a proof of concept showing how a given bug could be exploited, a plain explanation of the issue, and, where one exists, a suggested fix. That is a meaningfully more complete output than a typical static analysis tool, which usually flags a pattern and leaves the exploitability and the fix entirely to you.

The closest comparison Anthropic itself draws is Google's OSS Fuzz, the fuzzing infrastructure Google has run against open source projects for years. The mechanism is different, OSS Scanner uses model reasoning over a cloned and built repository rather than randomized input generation, but the goal and the opt in structure are the same: a major platform offers continuous, free security tooling to projects that would otherwise have to pay for it or go without.

The part worth sitting with: no human review

Anthropic is explicit that OSS Scanner reports are fully model generated and are not reviewed by a person before a maintainer sees them. The stated reasoning is straightforward: requiring human review before every report would cap how fast and how often scans could run, which defeats the purpose of continuous coverage. The tradeoff is that some reports will contain errors, including incorrect severity ratings, and Anthropic says so directly rather than implying otherwise.

Anthropic's target is a true positive rate above 90 percent. That is a strong number for an automated tool, and it also means that for every ten reports a maintainer receives, close to one, on average, will not be a real issue. For a solo maintainer or a small team, that is a very different workload than a single security email a year. It is a steady stream that needs its own triage discipline, and teams that already run a vulnerability disclosure process should plan for OSS Scanner output to need the same kind of first pass filtering they apply to any other automated scanner, not a free pass to merge suggested fixes unreviewed.

How a project actually enrolls

Enrollment is a pull request to the anthropics/oss-scanner GitHub repository containing a YAML configuration file. The required fields are narrow by design:

repository: https://github.com/your-org/your-project
contact_email: security@your-project.org
dockerfile_path: ./docker/scan.Dockerfile

The Dockerfile has to set up the environment, install dependencies, and build the project so an offline agent can audit it without network access during the scan itself. Optional fields let a maintainer add CC email addresses, a project homepage, a GPG public key so report emails arrive encrypted, a path to a threat_model.md file that gives the scanner context on what matters most to protect, and a disabled: true flag to pause reports without pulling out of the program entirely. Anthropic says it selects enrolled projects using criteria similar to Google's OSS Fuzz, meaning widely depended upon software is the priority, not every repository that applies.

Disclosure timing is different from a typical bug bounty

Most coordinated disclosure programs start a clock, often 90 days, the moment a report goes out, whether or not anyone has confirmed the bug is real. OSS Scanner does not default to that. Because reports are unreviewed and may include false positives, Anthropic does not impose a 90 day window automatically. If a finding is later validated by hand through Anthropic's existing coordinated vulnerability disclosure program, the 90 day clock starts then, from the moment the maintainer is notified of that validation, not from the original automated report. That is a meaningful difference if you are a maintainer trying to plan a patch release around a disclosure deadline: the automated report alone does not start that deadline.

Two programs, one mission, worth not confusing

It is easy to conflate OSS Scanner with the Critical Infrastructure Defense Program, since both launched under the Cyber Mission umbrella on the same day. They are distinct. CIDP is built around eleven named partners, Accenture, Booz Allen, CrowdStrike, Deloitte, Dragos, Hitachi, Insane Cyber, Nozomi Networks, Palo Alto Networks, PwC, and Rockwell Automation, and is aimed at power grids and other critical infrastructure operators, not general open source maintainers. OSS Scanner is the piece that is directly actionable for a software team today: you enroll a repository, you do not need a partnership agreement.

When this applies to your stack

If your organization maintains a public repository with meaningful adoption, or you depend heavily on a smaller open source project that has never had a real security audit, OSS Scanner is worth evaluating this quarter, specifically the triage capacity question. Getting real value out of free, continuous, AI generated vulnerability reports means having someone who can look at each one quickly enough to act on genuine findings without drowning in the roughly one in ten that will not hold up. Building that intake process, and deciding where AI generated security findings fit inside an existing vulnerability management pipeline, is exactly the kind of infrastructure and security tooling work we help engineering teams get ahead of before the backlog becomes the emergency.

FAQ

Is Anthropic's OSS Scanner actually free? Yes. There is no cost to enroll a project, and Anthropic states there are currently no plans to charge for it.

Do I need to trust every report OSS Scanner sends? No. Reports are generated by Anthropic's models without human review, and Anthropic expects a true positive rate above 90 percent, which means a meaningful share of reports will need to be dismissed after your own triage.

How is this different from Google's OSS Fuzz? Both offer free, continuous security scanning to enrolled open source projects, but OSS Fuzz finds bugs through automated fuzzing, feeding randomized inputs into a program, while OSS Scanner uses Anthropic's language models to reason about the code directly and produce a proof of concept and suggested fix for each finding.

When does the 90 day disclosure clock start for an OSS Scanner finding? Not automatically when the report is generated. It starts only if the finding is later validated by hand through Anthropic's coordinated vulnerability disclosure program, counted from when the maintainer is notified of that validation.