On October 8, 2026, Anthropic launched the Anthropic Cyber Mission, a long-term security initiative that puts frontier AI models to work defending the systems the public relies on. The short version: Anthropic will give free AI-powered vulnerability scans to open-source projects through a service called OSS Scanner, and it is backing a Critical Infrastructure Defense Program with 11 founding partners to help secure power grids, water systems, transport networks, and industrial systems. The company is contributing AI models, engineering support, threat research, and funding to both efforts.
What Is the Anthropic Cyber Mission?
The Cyber Mission is Anthropic’s umbrella program for defensive cybersecurity work. Rather than a single product launch, the company describes it as a long-term, collaborative effort that will grow through partner learning, new tools, research, and resources. Its initial work splits across two fronts: critical infrastructure defense and open-source software security. A third pillar is expanded access for cyber defenders through Anthropic’s Cyber Verification Program, which now absorbs the earlier Project Glasswing pilot.
The framing behind the announcement is blunt. Anthropic argues that frontier AI models can be misused to find vulnerabilities and run cyber operations, and that state-sponsored attackers have spent years establishing footholds across critical sectors so they can disrupt them later. Defenders of infrastructure and the open-source community have deep security experience, the company says, but they face severe resource shortages. The mission is meant to close that gap.
The Critical Infrastructure Defense Program
The Critical Infrastructure Defense Program, or CIDP, targets the service providers that help infrastructure operators secure operational technology, the OT systems running power, water, manufacturing, and transport networks. The program provides participating organizations with access to Anthropic’s AI models, engineering support, and threat research.
Eleven founding partners back the program at launch. Confirmed names so far include CrowdStrike, Accenture, Deloitte, and Palo Alto Networks. CrowdStrike says it has partnered with Anthropic on an offering that combines its own security tools with Anthropic’s AI models.
OT security is a fitting focus. Many industrial systems cannot be taken offline for patching the way a laptop can, which makes verifying and fixing a vulnerability far more burdensome. AI assistance with triage, analysis, and remediation planning is genuinely useful there, provided humans stay in the loop.
OSS Scanner: Free AI Vulnerability Scans for Open Source
The headline tool for most developers is OSS Scanner, a free, opt-in service that runs regular security scans of eligible open-source software using Anthropic’s strongest models. Anthropic projects a true-positive rate above 90 percent for the scanner, which would put it well ahead of traditional automated scanning tools on accuracy.
How OSS Scanner Works
Maintainers apply through the OSS Scanner GitHub repository using a published enrollment template. Once accepted, the service traces how data moves through the codebase, flags weaknesses across interconnected components, and can generate draft patches for human review. Anthropic says each finding passes through a multi-stage verification pipeline designed to cut down the false-positive noise that has plagued automated scanners for years.
One important caveat: the reports delivered to maintainers are AI-generated without prior human review, and Anthropic itself notes they may contain inaccuracies. The scanner does the heavy lifting of discovery, but the final triage decision still belongs to the project team.
Who Is Eligible
Eligibility decisions weigh whether a project has what Anthropic calls a critical impact on infrastructure and user security, along with the team’s ability to review and act on reported vulnerabilities. That bar makes sense: a scanner is only useful if someone on the other end can process the findings. Widely depended-upon libraries and tools maintained by small volunteer teams are exactly the profile the program is built for.
This matters beyond the maintainers themselves. Open-source code underpins a huge share of the world’s applications, and supply-chain attacks have shown how one compromised package can ripple outward. The recent Shai-Hulud worm, which stole developer credentials through an npm compromise, is a reminder of why better automated detection in the open-source ecosystem benefits everyone downstream.
Project Glasswing: The Numbers Behind the Mission
The Cyber Mission did not appear out of nowhere. Anthropic has been building toward it most of the year. In May it launched Project Glasswing, which paired an unreleased model with AWS, Google, Microsoft, Cisco, CrowdStrike, JPMorganChase, and the Linux Foundation to hunt bugs across critical open-source infrastructure. According to published reports, Glasswing found more than 10,000 high- or critical-severity vulnerabilities in its first month, and its cumulative tally has since passed 23,000 findings across more than 1,000 projects, touching software as widely used as NGINX, jq, Mastodon, ImageMagick, FreeRDP, and MinIO. Anthropic says it confirmed 90.6 percent of what the model flagged.
Earlier this week, Anthropic folded Project Glasswing into its Cyber Verification Program, broadening eligible defenders’ access to its models. The Cyber Mission is the next step down from Glasswing’s institutional scale: self-serve access for ordinary open-source maintainers, not just the large foundations invited into the pilot.
Why Anthropic Is Doing This Now
The timing reflects a shift both sides of the security industry are feeling. AI is increasingly the weapon attackers reach for, so defenders want the same firepower. AI agents have already demonstrated offensive capability in the wild: an AI agent recently compromised a cybersecurity nonprofit in seconds through a Zammad zero-day, and researchers keep finding new prompt-injection and agent-abuse techniques. At the same time, defensive AI is posting real results: Google’s PageBreak AI agent found more than 500 XSS vulnerabilities by working like a human security researcher.
Surveys back up the urgency. A PwC survey of nearly 4,000 executives reported this week found that less than a quarter of security, technology, and business leaders feel comfortable deploying AI agents without human controls, with many CISOs worried that agents inherit every permission misconfiguration and operate at machine speed. Programs like the Cyber Mission are, in part, an attempt to build the supervised, defender-side tooling that makes that discomfort manageable.
The Criticism: Noise, Triage Burden, and Unreviewed Reports
Not everyone is cheering. Google has cautioned that automated security findings can create extra triage work for maintainers, and discussions in the Open Source Security Foundation have warned that AI-generated disclosures can add low-quality noise to already-stretched volunteer teams. Those concerns are legitimate: a scanner with a 90 percent true-positive rate still produces one false alarm in ten, and each one costs a maintainer time.
Anthropic’s answer is the multi-stage verification pipeline and the eligibility bar requiring teams that can handle the reports. Whether that is enough will only be clear once the service is running at scale. For now, maintainers should treat OSS Scanner as a powerful first pass, not a replacement for human review.
What This Means for Developers and Security Teams
If you maintain an open-source project with meaningful downstream impact, the practical move is straightforward: review the enrollment template in the OSS Scanner GitHub repository and apply. Free, model-grade scanning with draft patches is the kind of resource small teams rarely get. If you work in critical infrastructure or OT security, watch what the founding partners ship; the CrowdStrike integration will be the first concrete product to judge.
The bigger picture is that AI-assisted defense is moving from research demos to production programs. Anthropic’s Cyber Mission, Google’s PageBreak, and the growing crop of agentic security tools all point the same direction: the teams that learn to supervise AI security tooling well will pull ahead of those that treat it as either magic or noise.
Frequently Asked Questions
What is the Anthropic Cyber Mission?
It is a long-term security initiative Anthropic announced on October 8, 2026, to help defend critical infrastructure and open-source software. It combines AI models, engineering support, threat research, funding, and industry partnerships, starting with the Critical Infrastructure Defense Program and the free OSS Scanner service.
Is OSS Scanner really free?
Yes. OSS Scanner is a free, opt-in service for approved open-source projects. Maintainers apply through the OSS Scanner GitHub repository using the published enrollment template, and accepted projects receive regular AI-powered security scans at no cost.
Which open-source projects qualify for OSS Scanner?
Anthropic weighs whether a project has a critical impact on infrastructure and user security, along with the maintaining team’s ability to review and address reported vulnerabilities. Projects with wide downstream adoption are the intended beneficiaries.
What was Project Glasswing?
Project Glasswing was Anthropic’s earlier pilot that paired an unreleased AI model with partners including AWS, Google, Microsoft, Cisco, CrowdStrike, JPMorganChase, and the Linux Foundation to hunt vulnerabilities in critical open-source infrastructure. Reports credit it with more than 23,000 findings across 1,000-plus projects, with Anthropic confirming 90.6 percent of flagged issues. It has now been folded into the Cyber Verification Program.
Who are the partners in the Critical Infrastructure Defense Program?
The program launched with 11 founding partners. Confirmed names include CrowdStrike, Accenture, Deloitte, and Palo Alto Networks. CrowdStrike is building an offering that combines its security tools with Anthropic’s AI models.