Anthropic Launches Critical Infrastructure Defense Program
Anthropic has launched a critical infrastructure defense program, entrusting its cutting-edge Claude model, on-site engineers, and threat research to vendors already responsible for protection. The initial focus includes operational technologies behind the power grid, water supply, factories, and transportation networks, as well as government systems. The initiative is placed under the newly established Anthropic Cyber Mission.
There are 11 founding partners. Consulting partners include Accenture, Booz Allen, Deloitte, and PwC. Security partners are CrowdStrike, Palo Alto Networks, and industrial security firms Dragos, Insane Cyber, and Nozomi Networks. Equipment partners are Hitachi and Rockwell Automation. Anthropic states that several partners are already using Claude to fix vulnerabilities and assist clients in doing the same. The company clarifies that AI cannot resolve all challenges in critical infrastructure; the goal is to identify and patch weaknesses before power outages or water quality issues arise.
The controllers, control software, and industrial networks of these systems are designed for decades of lifespan, many cannot be patched while offline, and known defects can persist for years. Operators, regardless of size, rely on a few trusted vendors to determine where vulnerabilities lie and what patches can be applied to operational systems. The plan is to first learn from this small group of vendors which practices are effective and executable. Reports will be delivered faster but may contain inaccuracies. The long-term goal is to automate most sorting and patching and develop new security architectures and coding standards.
The same mission has also launched a free scanning service for open-source projects called OSS Scanner, using the company's strongest model for regular vulnerability scans. Previously, Anthropic committed $100 million in usage credits through Project Glasswing and donated $4 million to organizations including OpenSSF and the Apache Software Foundation. The company also claims to have provided cutting-edge models and incident response, red team support to over half of U.S. states and large critical infrastructure operators.
The plan does not directly hand models to every water plant or substation. Access is granted to companies managing security projects for operators, safeguarding industrial networks, and manufacturing and patching equipment. The models come with usage controls to prevent them from being used to attack these systems. Several partners have already begun remediation, not just signing a list.
Buyers are security and industrial control vendors holding clients in the power grid, water, and transportation sectors. Sellers are Anthropic's cutting-edge model credits and on-site engineers. The driving force behind incidents comes from AI lowering the cost and speeding up attacks, along with the inability of old systems to go offline. Funding flows from usage credits and engineering support to partner client projects. The 11 partners and open-source maintainers benefit. Smaller operators without on-site slots and vulnerability services still priced by manual sorting are under pressure. Reports will be faster, but errors will also be included.
Source: Public Information
ABAB AI Insight
This initiative follows Anthropic's Project Glasswing. Glasswing initially opened the strongest model credits to vetted defenders, committing $100 million in usage credits and donating $4 million to OpenSSF and the Apache Software Foundation. The company has also provided scanning, patching, incident response, and red team support to over half of U.S. states and large facility operators. The critical infrastructure defense program consolidates the same set of credits from scattered pilots into 11 named partners, expanding the target from state governments to the industrial control supply chain.
Funds and personnel do not enter the water plant's accounts, but rather advanced partners' delivery contracts. Accenture, Deloitte, PwC, and Booz Allen sell security projects. CrowdStrike and Palo Alto Networks sell network protection. Dragos, Nozomi Networks, and Insane Cyber sell industrial network monitoring. Hitachi and Rockwell Automation sell controllers and patches. Claude credits and on-site engineers are embedded in these contracts, allowing Anthropic to obtain usage data and defect samples from critical infrastructure scenarios. OSS Scanner exchanges free scans for vulnerability distributions in open-source code.
This can be compared to Microsoft's prioritization of security across the company after 2023, embedding Copilot into Defender, and Google's Project Zero and open-source vulnerability rewards for early disclosures. The difference is that Anthropic does not have its own industrial control devices and must rely on Rockwell and Dragos for on-site access. The industry position is at the control stage: the model's ability to find vulnerabilities is already public, and competition has shifted from whose model is stronger to who can access and patch operational systems. OpenAI is similarly providing strong models to defenders, indicating that the credit battle has begun.
Structurally, this represents a reconstruction of the industrial chain. Operational technology was previously segmented by equipment vendors, integrators, and security vendors, with patching rights confined to downtime windows. Cutting-edge models harvest callable credits for finding vulnerabilities and writing patches, with on-site engineers deciding whether they can be applied to operational controllers. The mechanism is that channels are scarcer than models: water plants do not procure basic models but rather purchase repair recommendations from trusted vendors. Those who enter the delivery processes of these 11 partners gain access to defect samples that others cannot reach. Reports are allowed to contain errors, but the responsibility still lies with partners willing to act on operational systems.
ABAB News · Cognitive Law
Models find vulnerabilities; downtime windows determine whether patches can be applied.
Free scans exchange for defect samples, not market share.
Models that cannot access sites are merely another report.