Table of Contents
Anthropic has launched the Anthropic Cyber Mission, a long-term effort to apply frontier Claude models to defensive cybersecurity work across critical infrastructure and open-source software.
The program starts with two initiatives. The first, the Critical Infrastructure Defense Program, will bring Claude models, on-site engineers, and Anthropic threat research to companies that help secure operational technology used in power grids, water systems, transportation networks, and industrial environments. The second, OSS Scanner, offers free opt-in vulnerability scanning for critical open-source projects using Anthropic's strongest models.
Anthropic framed the launch as a response to a widening gap in cyber defense. AI systems are making vulnerability discovery cheaper and faster, while verification, disclosure, patching, and deployment still depend heavily on human maintainers and security teams.
Critical Infrastructure Defense Starts With 11 Partners
Anthropic's Critical Infrastructure Defense Program is aimed at security vendors, manufacturers, and technology providers that already work with operators of industrial and public infrastructure. Its 11 founding partners are Accenture, Booz Allen, CrowdStrike, Deloitte, Dragos, Hitachi, Insane Cyber, Nozomi Networks, Palo Alto Networks, PwC, and Rockwell Automation.
The focus is operational technology, or OT, the control systems, industrial equipment, and production networks that run physical infrastructure. These systems are difficult to defend because they often operate for decades, cannot always be taken offline for patching, and may combine legacy software with real-time safety requirements.
Anthropic said several partners are already working with Claude to fix vulnerabilities and help customers do the same. Instead of replacing industrial security expertise, the company gives existing providers access to frontier models, on-site technical support, and threat research they can apply inside defensive workflows.
The partner-led model suits this work because critical infrastructure security is rarely a simple software update problem. A vulnerability in an enterprise web application can often be patched with standard release processes. A vulnerability in an industrial control system may require coordination across plant operators, equipment manufacturers, safety teams, regulators, and vendors that understand the local environment.
Sam Rubin, senior vice president of Unit 42 at Palo Alto Networks, said in Anthropic's announcement that threat actors are using AI "to find and exploit exposures at machine speed." Dragos CTO Jon Lavender said the program helps align frontier AI with a defense strategy built around the realities of OT.
The effort builds on Anthropic's state, local, tribal, and territorial cyber program, which launched in June. Anthropic said that program has offered Claude and technical support to more than half of U.S. states, including work on code scanning, patching, incident response, red teaming, and other security workflows.
OSS Scanner Sends Model-Generated Reports to Maintainers
OSS Scanner addresses a different security bottleneck: the volume of vulnerabilities AI systems can now surface in open-source code.
According to Anthropic's technical post on the scanner, the company's models found more than 29,000 candidate vulnerabilities over the past six months while scanning important open-source projects. Human reviewers manually triaged about 6,000 of them. In some cases, maintainers asked Anthropic to send all available reports, even without human validation, and the company said it has sent nearly 5,000 unverified reports directly to maintainers that requested them.
OSS Scanner turns that experience into an opt-in service. Enrolled projects receive periodic scans from Anthropic's strongest models, including Claude Mythos, at no cost. The service is inspired by Google's OSS-Fuzz, but instead of relying on traditional fuzzing alone, it uses language models to analyze code and generate vulnerability reports.
Each report can include a proof of concept, an explanation of the issue, a bisection showing when the bug was introduced where possible, and a suggested patch. Anthropic said the reports are generated entirely by models and sent without human review or triage, which allows faster and more frequent scanning but also means projects need enough maintainer capacity to evaluate the findings.
Anthropic said expert penetration testers validated an early version of the scanner by reviewing 97 critical and high-severity findings across 48 projects. Of those, 85 findings, or 88%, met the bar for Anthropic's coordinated vulnerability disclosure process. Of the remaining 12, 11 were real issues that duplicated known problems or other scan findings, while one was invalid.
Anthropic expects the scanner's true-positive rate to exceed 90%. The team acknowledged that severity ratings can be inflated or that the model may misunderstand a project's threat model, which is one reason the service is aimed at projects with the ability to triage incoming reports.
Early maintainer feedback in Anthropic's post suggests the reports have already reached production projects. Todd Ouska of wolfSSL said the project received 74 reports, all but two of which were valid, and that five became CVEs. curl creator Daniel Stenberg said OSS Scanner helped find multiple issues in curl, including "one of the worst curl vulnerabilities reported in the last few years."
The AI Security Tradeoff
There's a great deal of tension in AI security today because of the inherent tradeoff.
The same models that can help defenders identify flaws in code, automate repetitive security work, and draft patches can also lower the cost of offensive vulnerability discovery.
Anthropic cited CyberGym benchmark results showing that LLMs improved from finding less than 20% of vulnerabilities at the start of last year to more than 85% this year. That progress helps explain why the company is moving from research projects such as Project Glasswing into production-facing services for maintainers and infrastructure defenders.
Open-source software is especially exposed because a small set of widely used libraries, developer tools, and infrastructure projects supports much of the modern software stack. AI companies, cloud providers, robotics teams, crypto networks, and decentralized AI projects all depend on shared open-source components. Free vulnerability scanning for critical projects can therefore function as defensive infrastructure beyond the individual repositories that enroll.
OSS Scanner also runs as an autonomous agent. Anthropic says the discovery process, proof of concept, explanation, and suggested remediation are all model-produced end to end, not AI summaries of human research. That makes it a live example of autonomous security agents performing real work on public codebases, with humans still responsible for judgment, validation, and patch acceptance.
Anthropic's own forecast is measured. The company wrote that AI may favor attackers in the near term because exploitation costs are falling while verification and repair remain slow. Its stated goal is for AI to favor defense within two years.
For now, the Cyber Mission gives Anthropic two concrete channels for pushing its models into defensive use. One runs through the companies that secure physical infrastructure, and another through the maintainers responsible for widely used open-source code. Companies building critical infrastructure security products can register interest in CIDP, while core maintainers of critical open-source projects can enroll in OSS Scanner.
