On October 8, 2026, Anthropic launched Anthropic OSS Scanner, a free opt-in service that runs periodic AI vulnerability scans on eligible open-source projects. The scans use Anthropic's strongest models, including Claude Mythos, and send reports as soon as they are ready.

The big trade-off is speed versus review. Reports are fully model-generated. There is no human triage on this fast track. Anthropic will keep its normal coordinated vulnerability disclosure (CVD) process for verified reports, especially for projects that cannot triage raw AI findings on their own.

I am Nitin from Debugged Pro. We help agencies and product teams in India and the USA ship software that depends heavily on open source. Below is a plain briefing on what shipped, who it is for, and what you should do next.

Anthropic Project Glasswing and security research visual for OSS Scanner context — Anthropic OSS Scanner builds on Project Glasswing work using Claude to find vulnerabilities in critical software. Image via Anthropic.

What Anthropic OSS Scanner Is (and Is Not)

Anthropic OSS Scanner is an opt-in vulnerability scanner for open-source maintainers. Projects that join get thorough, periodic security scans at no cost. Anthropic says the service is informed by its work using Claude during Project Glasswing.

It is not a free enterprise scan of every private app. Claude Security remains the commercial product for companies that want code scanning and patching inside their own lifecycle. OSS Scanner is the public-goods path for critical open source.

Anthropic also points to Google's OSS-Fuzz as inspiration. OSS-Fuzz uses fuzzers. Anthropic OSS Scanner uses language models. Both aim to catch bugs earlier in software that many other systems depend on.

For the primary announcement, read Anthropic's launch post on the opt-in vulnerability-finding service. For a second news angle, see The Verge coverage of Anthropic's free AI security scans by Stevie Bonifield.

Why Anthropic Built OSS Scanner Now

Anthropic's post lays out a clear bottleneck. Over the last six months, its latest models found more than 29,000 candidate vulnerabilities in important software projects. Staff were able to manually review and triage about 6,000 of those.

That gap matters. Anthropic says exploits can now be built in minutes. Projects that find and fix issues faster reduce the window attackers can use. Maintainers who already trusted Anthropic's early reports also asked for bulk dumps of unverified findings and patches. Anthropic says it has sent nearly 5,000 reports that way after maintainers asked for everything on hand.

Model quality on vulnerability discovery also moved fast. On CyberGym, an academic vulnerability-finding benchmark Anthropic cites, LLMs went from finding under 20% of vulnerabilities early last year to finding over 85% this year. Maintainers are no longer only getting low-value AI spam. They are getting reports that can include a real exploit and a candidate fix.

Anthropic OSS Scanner is the optional fast track for teams that want those reports before a human finishes review.

How Anthropic OSS Scanner Reports Work

Each report is meant to be actionable. Anthropic says early disclosures included:

A self-contained reproducer
An explanation of the vulnerability
A bisection to when the bug was introduced, where possible
A candidate patch when one is available

Early pipeline tests with dozens of projects produced hundreds of bug reports. Some findings could be chained into unauthenticated remote code execution exploits against those projects. That is exactly why maintainers need a triage lane before they merge AI-suggested patches.

Screenshot style visual related to Anthropic open-source security scanning coverage — News coverage of Anthropic's free AI security scans for open-source projects. Image via The Verge.

Early quality signals (verified numbers only)

Anthropic asked expert penetration testers who review CVD findings to check 97 critical and high-severity scanner results across 48 projects. Of those, 85 (88%) met the bar for Anthropic's CVD process. Of the remaining 12, 11 were real but duplicated known issues or other scan findings. Only one was invalid.

Maintainer quotes in the launch post are also useful context:

PostgreSQL: Noah Misch said an unusually high fraction of findings uncovered real defects, and some fixes were usable nearly as-is.
OpenSSL Corporation: Anton Arapov said early AI reports from about 18 months ago were poor, but Anthropic's later reports, including raw model output, were as good and sometimes better than human reports, especially with a real exploit attached.
wolfSSL: Todd Ouska said 74 reports arrived; all but two were valid; five became CVEs.
HotCRP: Eddie Kohler praised thorough reports and strong understanding of a complex permission model.

Anthropic still warns that severity ratings can be inflated and that the scanner can misunderstand a project's threat model. Treat every report as a lead, not as a merge request.

Enrollment: Who Can Join and How

Eligibility follows a similar idea to OSS-Fuzz. Anthropic wants established projects with a critical impact on infrastructure and user security. Decisions are case by case. Factors include exposure to remote attacks and how many users or other projects depend on the code.

Core maintainers enroll by opening a pull request on the anthropics/oss-scanner GitHub repository. The config lives at projects/<name>/project.yaml. Required fields include:

repo: git URL to clone, optionally with a branch
primary_contact: email for reports and out-of-band contact
Dockerfile: path to a Dockerfile that installs dependencies and builds the project so a fully offline agent can audit it

Optional fields include extra CC emails, homepage, a threat model file path, a GPG key for encrypted mail, and a disabled flag to pause reports. Anthropic manually checks that the submitter is a core maintainer before enrollment.

The Dockerfile is important. It is built with network access. Everything after that runs with internet access disabled inside hardened sandboxes. Anthropic recommends confirming tests pass inside the built container before you open the PR.

A threat model file helps a lot. You can document which inputs are adversarial, what is out of scope, how you want severity labeled, and how patches should look. Without it, the scanner will guess.

Disclosure Policy for Unvalidated Findings

Anthropic will not put a 90-day coordinated disclosure clock on these unvalidated OSS Scanner findings. The company says it does not want to force maintainers to race through reports that humans have not yet verified.

If Anthropic later validates a report through CVD, a 90-day clock can start from the date the maintainer is notified of that human validation. Anthropic also says it may later impose a disclosure period on some high-severity reports as confidence grows, with notice and an opt-out.

Projects can pause with disabled: true or leave by deleting their project directory via PR. After opting out, they fall back to the standard CVD path only.

Anthropic OSS Scanner vs Claude Security vs CVD

Agencies should keep these three lanes separate in client conversations.

Lane	Who it serves	Human review	Cost
Anthropic OSS Scanner	Eligible open-source projects	No on the fast track	Free
Anthropic CVD	Projects that need verified reports	Yes	Free (Anthropic process)
Claude Security	Enterprise codebases	Product workflow for teams	Commercial

Claude for OSS is a related support offer: free Claude Max 20x subscriptions to help maintainers remediate issues. Anthropic's Cyber Verification Program is aimed at qualifying security professionals who need advanced cyber capabilities and reduced blocking classifiers.

If your client sells a commercial SaaS product, do not tell them OSS Scanner will scan their private monorepo for free. Point them to Claude Security or another commercial scanner, and use OSS Scanner only for the public upstream projects they maintain or heavily depend on.

What This Means for Agencies in India and the USA

Most agency stacks sit on Linux, OpenSSL, PostgreSQL, language runtimes, CI plugins, and dozens of smaller libraries. When those projects patch faster, your client risk drops. When they get flooded with AI reports they cannot triage, your clients inherit delay and noise.

Here is the practical split we recommend for India and USA teams:

If you maintain critical OSS: decide whether you want the CVD lane, the OSS Scanner fast track, or both over time.
If you only consume OSS: track which dependencies enroll, watch advisory feeds, and budget patch windows when high-signal AI reports turn into CVEs.
If you sell security or platform work: offer a triage retainer for AI-generated vulnerability reports, with clear SLAs for reproduce, severity, and patch review.

Also keep policy and privacy context in view. Our earlier note on the Anthropic usage policy 2026 changes for agencies covers high-risk use rules that still apply to how teams use Claude. For client data handling around agentic tools, see our briefing on AI data protection and the ICO's agent scrutiny.

Six-step agency checklist for adopting Anthropic OSS Scanner responsibly — A practical Anthropic OSS Scanner checklist for agencies in India and the USA. Graphic: Debugged Pro.

Agency Action Checklist for Anthropic OSS Scanner

Use this as a one-week pilot plan, not a press-release checklist.

Inventory critical OSS you maintain or heavily fork. Mark anything that processes untrusted input or sits on a remote attack path.
Pick the disclosure lane. If your maintainer team is small, stay on CVD for now. If you already clear verified high and critical bugs quickly, consider OSS Scanner.
Assign one owner for enrollment. That person opens the GitHub PR, owns the Dockerfile, and routes email reports into your tracker.
Write the threat model file first. Spell out scope, severity rules, and patch style before the first scan lands.
Treat reports like bug-bounty intake. Reproduce first. Deduplicate second. Patch third. Never merge a candidate patch only because the model wrote it.
Credit report IDs when you patch. Anthropic asks for a line such as "Discovered by Anthropic's OSS Scanner, as vulnerability ANT-2026-ABCD1234." That helps track real impact.
Brief client security leads. Explain that free AI scans on upstream OSS do not replace commercial scanning of private code.
Measure noise. After the first bundle, record true positives, duplicates, severity disagreements, and time-to-reproduce. Decide whether to stay enrolled.

If your team needs help wiring this into a wider delivery process, Debugged Pro's software development consulting practice can map OSS risk, enrollment, and triage into the same sprint cadence you already run for product work.

Key Takeaways

Anthropic OSS Scanner launched on October 8, 2026 as a free opt-in scanner for eligible open-source projects.
Scans use Anthropic's strongest models, including Claude Mythos, and reports arrive without human review.
Anthropic remains bottlenecked on human triage of findings (about 6,000 reviewed out of more than 29,000 candidates over six months).
Early validation on 97 critical/high findings showed 88% meeting CVD standards, with only one clear false positive in that sample.
Enrollment is a GitHub PR with repo, contact, and Dockerfile details; eligibility mirrors OSS-Fuzz style critical-infrastructure criteria.
Agencies should separate OSS Scanner, CVD, and Claude Security when advising clients in India and the USA.

Final Notes for Maintainer and Agency Teams

Anthropic OSS Scanner is a serious tool for projects that already run a mature security process. It is not a magic inbox cleaner for every repository on GitHub. The value shows up when you can reproduce findings quickly, reject bad severity calls, and ship patches before attackers reuse the same lead.

If you maintain critical open source, read the FAQ on Anthropic's OSS Scanner page, prepare the Dockerfile and threat model, and only then enroll. If you mainly consume open source, watch the projects you depend on and plan patch windows. If you sell agency services, sell triage and verification, not blind trust in model output.

About the author: Nitin Kumar is Lead Project Coordinator at Debugged Pro. He works with product and agency teams in India and the USA on practical AI and software delivery.