The 10 Things Anthropic's Free Security Scanner Is Actually Doing in 2026 (It Is Not Charity)

Everyone is arguing about AI taking jobs. Anthropic just volunteered its best models to guard power grids and patch your open-source repo for free.

Anthropic's OSS Scanner points frontier models at open-source codebases. Photo by Shahadat Rahman on Unsplash.

Let's stop pretending AI safety is a research paper. On October 8, 2026, Anthropic turned it into a security company.

Here is the uncomfortable truth: their models already found more than 29,000 candidate vulnerabilities in six months. They had no choice but to deal with them.

This article is not a product launch recap. It is about what happens when the company that built the weapon starts giving away the shield.

1. There is a new umbrella, and it has a dramatic name

Anthropic grouped its security work under one banner: the Anthropic Cyber Mission.

• Announced October 8, 2026, it covers two programs: the free OSS Scanner for open source and a critical infrastructure defense program.
• Earlier the same week, Anthropic folded Project Glasswing (which gave vetted organizations access to Claude Mythos since April 2026) into an expanded Cyber Verification Program.
• The stated lesson from Glasswing: finding vulnerabilities has never been easier, but verifying, prioritizing, and fixing them remains hard. And it "has not yet cut cyber risk by enough."

The so what: this is a strategy, not a stunt. (Source: SiliconANGLE)

2. Anthropic will scan your open-source project for free

The OSS Scanner is a free, opt-in, periodic vulnerability-scanning service for eligible open-source projects.

• It runs on Anthropic's strongest models, the frontier Claude family.
• It is funded by the Defender Advantage Fund, which Anthropic set up in August 2026.
• Anthropic also put money into the Python Software Foundation, the Apache Software Foundation, and Alpha-Omega and OpenSSF via the Linux Foundation.

The so what: the best model Anthropic has, pointed at your code, for zero dollars. (Source: CyberScoop)

Every report ships with a working exploit, not just a warning. Photo by Tobias Tullius on Unsplash.

3. Every report ships with a working exploit

This is the part that made me sit up. The scanner does not just flag problems.

• Each report includes a self-contained reproducer: a proof of concept showing the vulnerability actually works.
• Where the model can manage it, you also get a candidate patch plus a code-history bisection showing exactly when the flaw was introduced.
• Every finding comes with an explanation and suggested patching options, so the fix is essentially clear for an engineer.

The so what: they do not just tell you there is a hole. They hand you the weapon and the bandage. (Source: SiliconANGLE)

4. The reports skip human review, mistakes included

Reports go straight to maintainers with no human review, because speed matters more than polish.

• Anthropic admits some reports will contain mistakes, like wrong severity ratings.
• In a pre-launch accuracy test, expert pen testers checked 97 critical and high-severity findings across 48 projects and cleared 85 for disclosure.
• Of the remaining 12, all but one were real bugs that simply duplicated known issues or other scanners' findings.

The so what: speed over perfection is the entire design philosophy. (Source: SiliconANGLE)

5. The accuracy numbers are surprisingly good

Early trial results suggest this is not a demo dressed up as a program.

• Embedded encryption library maker wolfSSL received 74 reports during early trials. All but two were valid.
• Five of those wolfSSL reports became CVEs, the official cataloged vulnerabilities the industry tracks.
• Remember the scale here: these are findings in widely used software, generated by a model, with patches attached.

The so what: this is already producing real CVEs, not conference slides. (Source: SiliconANGLE)

The second half of the mission aims at operational technology: plants, substations, water systems. Photo by Kevin Ache on Unsplash.

6. This exists because Anthropic drowned in its own findings

The scanner was born from a backlog problem, which is the most honest origin story in tech.

• Over the past six months, Anthropic's models turned up more than 29,000 candidate vulnerabilities in widely used software.
• Human staff manually reviewed about 6,000 of them.
• Nearly 5,000 unreviewed reports, complete with proposed patches, were already sent to maintainers who asked for the full batch.
• The setup is openly inspired by Google's OSS-Fuzz.

The so what: the scanner is a backlog solution wearing a superhero cape. (Source: SiliconANGLE)

7. Enrollment is a GitHub pull request

Getting in is refreshingly un-corporate for a program this ambitious.

• Core maintainers enroll by submitting a pull request to an Anthropic GitHub repository.
• Eligibility follows the Google OSS-Fuzz test: critical impact on infrastructure and user security, decided case by case.
• Projects without the staff to handle raw findings still get human-verified reports through Anthropic's existing disclosure process.

The so what: open source, for open source, through open source. (Source: SiliconANGLE)

8. Eleven companies now guard the power grid with Claude

The second program is the critical infrastructure defense effort, and the partner list reads like a security conference attendee sheet.

• Founding partners (11): Accenture, Booz Allen Hamilton, Deloitte, PwC, Dragos, Insane Cyber, Nozomi Networks, CrowdStrike, Palo Alto Networks, Hitachi, and Rockwell Automation.
• Anthropic brings frontier Claude models, onsite engineers, and threat research to the firms protecting power grids and water systems.
• Several partners already use Claude to scan and patch code, advise which fixes are safe to apply to running systems, and help with incident response and red-teaming.

The so what: this is AI inside the substation, not AI in a slide deck. (Source: CyberScoop)

Operational technology is, in one executive's words, the next frontier for autonomous AI-enabled attacks. Photo by Markus Spiske on Unsplash.

9. More than half of U.S. states got the offer

This is not a pilot. The distribution is already nation-scale.

• Anthropic has offered frontier models and technical support to more than half of U.S. states and to large operators of critical infrastructure.
• The target is operational technology: plants, substations, and water systems where equipment must run for decades and often cannot be taken offline for patching.
• The long-term goal: automate most triage and patching, and develop new security architectures and coding standards.

The so what: when a lab offers its best models to half the states in the union, it is not a PR move. (Source: CyberScoop)

10. The catch nobody printed on the poster

I will not soften this, because the honesty rule cuts both ways.

• The commercial terms are undisclosed. As Axios noted, Anthropic has not said whether partners get free model access or who covers the compute costs.
• Patching industrial systems that cannot go offline remains an unsolved problem. A bad patch on a water plant is worse than the vulnerability.
• Anthropic's own verdict on its earlier effort: finding vulnerabilities has never been easier, but verification and repair is slow, people-dependent work, and Glasswing "has not yet cut cyber risk by enough."
• Anthropic expects AI to favor defenders within two years. Today, exploit costs keep falling.

The so what: free scans are the easy part, and even Anthropic admits it. (Source: SiliconANGLE)

How to start

If you maintain an open-source project: check whether your project meets the critical-impact bar, then submit the pull request to Anthropic's repo. Brace for raw, unreviewed reports. They will be fast and occasionally wrong.

If you are a security-curious developer: read the reproducer reports as free penetration-testing training. A working exploit with a bisection is a better teacher than any course.

If you just ship software: check whether your dependencies are enrolling. Free scanning only helps if maintainers actually apply the patches.

Do not sleep on this because it is free. Free is the strategy.

The real pattern

The labs spent two years building models that find bugs faster than any human can, and that created a surplus of vulnerabilities nobody could process. So now they are building the other half of the pipeline for free, because a world full of AI-found exploits with no AI-speed patching is a world where their own models become the problem. The real pattern is simple: whoever creates the surplus owns the cleanup.

Final thoughts

Free security from the people who made the insecurity. Take the scans. Patch fast.

And remember the uncomfortable bit: the same model that found the bug could have sold it.

Yes, AI helped write this. It did the typing. I gave the ideas and I did the editing and orchestrating.

If this saved you money or taught you something, buy me a coffee. It keeps the articles coming.

Follow along. Clap if it earned it.