Anthropic Cyber Mission: Free AI Scanner for Open Source

Anthropic is putting its models to work on defense. The company has announced Cyber Mission, a program it describes as long-term, with two goals: protecting critical infrastructure and protecting the open-source software that modern systems run on. The announcement has two parts. The first is a partnership program for infrastructure operators. The second is a free AI scanner that will check open-source projects for vulnerabilities on a regular basis.

Anthropic's reasoning is simple. Attackers already have access to powerful AI models, the company argues, and defenders still lack tools of comparable strength. Cyber Mission is its attempt to close part of that gap.

Two programs under one banner

1. The Critical Infrastructure Defense Program (CIDP)

The first part targets the organizations that keep physical systems running. CIDP is aimed at operators of power grids, water systems and transportation networks. Participants get access to three things:

• Claude models
• Anthropic engineers
• Threat analysis

The program launches with four founding partners: CrowdStrike, Palo Alto Networks, Deloitte and Rockwell Automation. The mix is worth noting. CrowdStrike and Palo Alto Networks are major cybersecurity vendors, Deloitte is a large consulting firm, and Rockwell Automation builds industrial automation technology. Together they cover security tooling, implementation support and the operational technology used in industrial settings.

Anthropic has been widening access to its security tools for some time. Its cyber verification program for defenders was recently opened to more users, and CIDP pushes that approach toward the sectors where a successful attack does the most physical damage.

2. The free "OSS" scanner

The second part is probably the more significant one for developers. Anthropic is offering a free AI scanner, called "OSS," that will check open-source projects at regular intervals. When it finds a vulnerability, it will:

• Flag the issue automatically
• Explain what the problem is
• Suggest a patch

The case for focusing on open source is a familiar one. Almost all modern software depends on open-source code, and much of that code is maintained by small teams of volunteers. These maintainers often look after components that large companies and public services rely on, but they rarely have the time or budget for steady, professional security review.

The accuracy question

Anthropic says it expects the scanner to reach an accuracy above 90 percent. That figure comes with an important caveat: the reports go out without human review, and Anthropic acknowledges they may contain errors.

For maintainers, that trade-off cuts both ways. A frequent automated check could catch flaws that would otherwise sit unnoticed for years. But a report that is wrong still takes time to read, reproduce and dismiss. For a volunteer team, the cost of checking false positives is real.

This is not a theoretical worry. Google recently paused an open-source bug bounty after a flood of AI-generated reports. Anthropic's model is different in one key way: the scanner is opt-in, not something that lands unasked in a maintainer's inbox.

Who can sign up

Participation is limited and voluntary. Maintainers of projects that are critical to infrastructure or to user safety can opt in through GitHub. Anthropic has not presented the scanner as a tool for every repository. It is aimed at code where a vulnerability would have serious consequences.

The opt-in design gives maintainers control. They choose whether they want automated reports at all, and they decide what to do with each suggested patch.

The defender's argument

The core claim behind Cyber Mission is that the balance between attack and defense has shifted. In Anthropic's view, offensive actors already use capable AI models, while the people defending systems lag behind. Reports of AI hacking tools used against Korean banks show why that argument is getting attention across the industry.

Anthropic's answer is to give defenders the same class of tools for free, or through structured partnerships, rather than leave them to work out access on their own.

Our Take

Cyber Mission suggests that AI labs increasingly see security tooling as part of their responsibility, not just a side market. The open-source scanner is the more interesting experiment. If it works, it could offer small volunteer teams a kind of regular review they could rarely afford before.

The weak point is the one Anthropic names itself: reports ship without human review. An accuracy above 90 percent sounds strong, but at scale even a small share of wrong reports could add work for exactly the people the program aims to help. The opt-in model helps limit that risk.

It is worth watching whether maintainers actually sign up, how often suggested patches are accepted, and whether Anthropic publishes data showing that real vulnerabilities were fixed. On the infrastructure side, the open question is whether CIDP produces measurable improvements for grid, water and transport operators, or stays mainly a partnership framework. Those results will show whether "AI for defenders" holds up in practice.