- On October 8, 2026, Anthropic launched the Cyber Mission: a Critical Infrastructure Defense Program with 11 founding partners and OSS Scanner, free and opt-in periodic vulnerability scans of open-source projects by its strongest models, including Claude Mythos.
- Maintainers enroll by opening a pull request to github.com/anthropics/oss-scanner with a project.yaml and a Dockerfile. Reports arrive by email, unreviewed by humans; in Anthropic's own validation, 85 of 97 critical and high findings (88%) met its disclosure bar.
- On October 6, Anthropic merged Project Glasswing into a Cyber Verification Program with Defense, Red Team and Specialized Access tiers. It says Glasswing partners found at least 129,000 verified vulnerabilities between April and July 2026, from partial data.
Contents
On October 8, 2026, Anthropic launched what it calls the Anthropic Cyber Mission, "a long-term commitment to securing the systems everyone depends on." It starts with a Critical Infrastructure Defense Program for companies that secure grids, water systems and other operational technology, and OSS Scanner, free, opt-in vulnerability scans of open-source projects by Anthropic's strongest models, including Claude Mythos. Two days earlier, on October 6, Anthropic folded Project Glasswing into an expanded Cyber Verification Program with three access tiers.
This story is based on Anthropic's announcements, the OSS Scanner repository and FAQ, and The Hacker News' coverage, checked October 12, 2026.
The short version
- OSS Scanner: free periodic scans for eligible open-source projects, enrolled by pull request. Reports are model-generated and unreviewed.
- Critical Infrastructure Defense Program: Claude models, engineers and threat research for 11 founding security and industrial partners.
- Cyber Verification Program (CVP): three tiers (Defense, Red Team, Specialized Access), each with Claude Opus 5.5, Sonnet 5.5 and Mythos 5.1, with fewer cyber blocks at each step up.
- Glasswing's tally: Anthropic says partners found "at least 129,000 verified software vulnerabilities between April and July 2026." That number is Anthropic's, from partial partner data.
How do open-source maintainers opt in to OSS Scanner?
According to the OSS Scanner repository and FAQ, enrollment works like Google's OSS-Fuzz:
- Check eligibility. Anthropic accepts "established projects that have a critical impact on infrastructure and user security," case by case, and will "manually validate you are a core maintainer."
- Open a pull request to
github.com/anthropics/oss-scannerthat adds one directory,projects/<name>/, containing aproject.yamland aDockerfile, plus an optionalthreat_model.md. - Fill in
project.yaml.repoandprimary_contactare required;auto_ccs,homepage,disabled,threat_modelandpgpare optional. Addresses in the file are public, so use a security alias. - Make the Dockerfile work offline, "so that a fully offline agent can conduct its security audit." Test it with the repo's
tools/validate.pyandtools/check <name>. - Wait for the scan. After merge, the project is built online and analyzed in an isolated VM with no internet access. Findings go by email to
primary_contact.
To pause reports, set disabled: true by pull request. To leave, submit a pull request that deletes your project directory. Anthropic doesn't give a fixed scan schedule; after the first scan it says it will "regularly scan projects."
The repository showed 186 pull requests on October 12, 2026. The Hacker News counted 116 when it reported the launch on October 9.
How accurate are the reports?
Each report includes a reproducer, an explanation and, where available, a candidate patch. Anthropic is upfront that they are "fully model-generated" and that "some will contain inaccuracies, such as a wrong severity rating."
Its two posts give slightly different figures, both vendor-reported. The Cyber Mission announcement says Anthropic "expect[s] a true-positive rate above 90%." The launch post says expert penetration testers checked 97 critical and high-severity findings across 48 projects: 85 (88%) met the bar for Anthropic's disclosure process, 11 were real but duplicates, and one was a false positive.
Because the reports aren't human-verified, Anthropic doesn't apply its usual 90-day disclosure deadline or publish them; if a human later validates a finding, the clock starts then. The reason for automating: over six months, Anthropic says, it found more than 29,000 candidate vulnerabilities in open-source code but manually triaged only about 6,000.
What is the Critical Infrastructure Defense Program?
The program brings "frontier Claude models, on-site engineers, and our threat research" to providers that secure operational technology and government systems. It starts with "a small cohort." The 11 founding partners are Accenture, Booz Allen, CrowdStrike, Deloitte, Dragos, Hitachi, Insane Cyber, Nozomi Networks, Palo Alto Networks, PwC and Rockwell Automation. Security vendors, system integrators and equipment manufacturers can register interest; the announcement doesn't describe applications from utilities themselves. Anthropic says some operational-technology fixes "might take decades" to apply safely, and that "in two years, AI will favor defense," but "in the near term, that may not be true."
What are the three Cyber Verification Program tiers?
The expanded CVP, announced October 6, merges Project Glasswing, which gave selected organizations access to Claude Mythos, with the original CVP, which reduced safeguards on Opus and Sonnet for vetted teams.
| Tier | For | Who can apply | Review time |
|---|---|---|---|
| Defense Access | SOC and incident response, malware reverse-engineering, vulnerability analysis | Security teams defending their own systems, critical infrastructure operators, smaller security firms, open-source maintainers, individual researchers with a disclosure track record | "Within a few days" |
| Red Team Access | Adds authorized penetration testing and red-teaming | Organizations only: in-house and government red teams, pen-testing firms | A few weeks (Defense Access meanwhile) |
| Specialized Access | Fewest cyber blocks, for testing safety systems such as flight systems, power grids and telecom networks | A limited set of organizations, each reviewed in depth with the US government; existing Glasswing members move here | Not stated |
Last verified: October 12, 2026 (Anthropic, "Expanding the Cyber Verification Program")
All tiers include Claude Opus 5.5, Sonnet 5.5, Mythos 5.1 and future models, and generally require data retention so Anthropic can monitor for misuse. CVP runs on the Claude Platform, Google Cloud's Vertex AI and Microsoft Foundry; on Amazon Bedrock only for customers eligible for Enterprise Frontier Safeguards, which Anthropic says is due "later this fall." Anthropic hasn't published CVP pricing; Mythos 5.1's list price is in our Claude model guide.
On Anthropic's own CyScenarioBench test (vendor-reported), Defense Access was blocked at some point in 46 of 50 trials; Red Team Access saw no blocks, and Opus 5.5 completed 34 of 50 tasks. Generally available models keep conservative safeguards that block most offensive cyber work, a choice our Opus 5.5 system card explainer traces to the model's cyber results.
Where does the 129,000 figure come from?
From Anthropic's October 6 announcement. It says Glasswing partners uncovered "at least 129,000 verified software vulnerabilities between April and July 2026," more than 33,000 of them rated critical or high, and that its own open-source scanning found another 5,500 verified vulnerabilities between April and October. The figures draw on partial data from 33 partner reports, and Anthropic says it expects "the true impact to be at least five times higher." No independent party has audited these counts, and Anthropic doesn't define "verified" beyond its partners' reports.
What's confirmed, reported and uncertain
- Confirmed (Anthropic): launch dates, the CIDP partners, OSS Scanner's enrollment steps, fields and opt-out, its validation result, CVP tiers, models, platforms and eligibility, and the Glasswing figures as Anthropic states them.
- Reported: The Hacker News' October 9 pull-request count.
- Uncertain: scan frequency, how many projects Anthropic will accept, CVP pricing, and independent confirmation of the 129,000 figure.
What it means for you
If you maintain a widely used project and can handle unverified high-severity reports, OSS Scanner costs a pull request and an offline Dockerfile. Security teams that hit Claude's cyber blocks should look at Defense Access first. Google is gating Gemini 4 Argon in a similar way, as our Gemini 4 Argon story explains. For how Mythos fits into Anthropic's business and its dealings with the government, see Anthropic explained.
About this storyBased on the sources linked below. Editorial standards
