The Critical Infrastructure Defense Program is starting with operational technology used in power grids, water systems and transportation networks, as well as government systems. Its 11 founding partners are Accenture, Booz Allen, CrowdStrike, Deloitte, Dragos, Hitachi, Insane Cyber, Nozomi Networks, Palo Alto Networks, PwC and Rockwell Automation.
These organizations span consulting firms, cybersecurity vendors and industrial equipment makers that already work with critical infrastructure operators. Anthropic said several of them are using Claude to identify and repair vulnerabilities and to support customers doing the same.
The initial phase will focus on a smaller group of providers while Anthropic evaluates which approaches are practical and effective. Companies that build security products or services for critical infrastructure can register interest as the program expands.
Anthropic is targeting operational technology because these systems can be unusually difficult to secure. Industrial controllers and related software are often designed to remain in service for many years, and taking them offline for maintenance can carry significant operational risk. In some cases, known vulnerabilities can remain unresolved for long periods because patches cannot be applied safely while equipment is running.
Anthropic said AI cannot solve all of those problems, but it believes frontier models can help defenders find weaknesses and develop fixes before those vulnerabilities are exploited.
The new program extends work Anthropic began earlier with state, local, tribal and territorial governments. The company said it has since provided frontier Claude models and technical support to more than half of U.S. states and to several large public critical infrastructure operators, covering tasks such as code scanning, patching, incident response and red teaming.
The second part of the Cyber Mission is OSS Scanner, a free service for open-source maintainers that want automated vulnerability reports from Anthropic’s models. The scanner was developed in response to maintainers who asked to receive findings directly, including results that had not gone through human review.
Each report can include a proof of concept, an explanation of the vulnerability and a suggested patch when one is available. Anthropic said the reports are generated entirely by models and sent directly to participating maintainers without manual screening.
That approach is intended to speed up delivery, but it also means some reports may include mistakes, including incorrect severity assessments. Anthropic said it expects more than 90% of the scanner’s findings to be true positives.
In testing, Anthropic asked expert penetration testers to evaluate 97 critical and high-severity findings across 48 projects. The company said 85 of those met the standard for its coordinated disclosure process. Eleven others were valid issues that duplicated known vulnerabilities or other reports, while one was determined to be a false positive.
Over the previous six months, Anthropic said its models identified more than 29,000 candidate vulnerabilities. About 6,000 were manually reviewed and triaged, while nearly 5,000 unverified findings were sent directly to maintainers that had asked to receive all available results.
OSS Scanner is intended for projects with enough capacity to evaluate a high volume of findings. Anthropic said projects without that capacity will continue receiving human-verified reports through its existing coordinated vulnerability disclosure process.
The company is also planning to extend the open-source program beyond vulnerability discovery. Its stated next steps include automating more of the triage and patching process and researching new approaches to writing and hardening software.
The Cyber Mission builds on Project Glasswing, an earlier effort launched with companies including Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, Microsoft, Nvidia and Palo Alto Networks. Anthropic previously committed up to $100 million in Claude usage credits and $4 million in direct donations through that initiative.
Project Glasswing has now been folded into Anthropic’s expanded Cyber Verification Program, with existing members moving into the program’s Specialized Access tier.
Anthropic said the Cyber Mission will continue expanding over the coming months, with more critical infrastructure partners and additional work around open-source software and supply-chain security. The company also said it plans to publish lessons from the program and collaborate with other AI developers, security companies and governments working on similar defensive efforts.
Anthropic’s broader forecast is that AI could increasingly favor defenders over the next two years by helping teams catch vulnerabilities earlier and build more secure software. The company cautioned, however, that the near-term environment may still favor attackers because exploiting flaws can happen faster than verifying and fixing them.
This analysis is based on reporting from Anthropic & unite.ai.
Image courtesy of Anthropic.
This article was generated with AI assistance and reviewed for accuracy and quality.
