The same week researchers revealed that an autonomous AI tool had breached seven South Korean banks with no human operator involved, Anthropic announced it was turning its frontier models on the problem from the opposite direction. The company’s new Cyber Mission, unveiled today, is its most ambitious security initiative yet: a structured effort to put Claude inside the organizations defending the power grids, water systems, transportation networks, and open-source software that modern civilization runs on.
This is not a press release dressed up as a program. It launched on October 8, 2026 with real partners, real code already running, and a free vulnerability scanner that has quietly found more than 29,000 candidate security flaws in widely used software over the past six months.
Two Programs, One Urgent Problem
The Cyber Mission has two arms. The first, the Critical Infrastructure Defense Program (CIDP), gives security companies that protect operational technology — the embedded systems inside power plants, water treatment facilities, factories, and substations — access to Anthropic’s frontier Claude models, on-site engineers, and the company’s internal threat research.
The eleven founding partners are a who’s-who of industrial cybersecurity: Accenture, Booz Allen Hamilton, CrowdStrike, Deloitte, Dragos, Hitachi, Insane Cyber, Nozomi Networks, Palo Alto Networks, PricewaterhouseCoopers, and Rockwell Automation.
Several of these companies were already quietly using Claude to find and fix vulnerabilities before today’s announcement. Now Anthropic is making the arrangement official, structured, and extensible to more sectors.
The second arm, OSS Scanner, is what separates the Cyber Mission from the usual category of corporate security announcements. It is a free, opt-in vulnerability scanning service for critical open-source software projects, modeled on Google’s long-running OSS-Fuzz program. Maintainers who enroll receive periodic scans from Anthropic’s strongest models. Each report comes with a self-contained proof-of-concept reproducer. Where possible, it also includes a candidate patch and a bisection of code history showing precisely when the flaw was introduced.
That last detail matters more than it sounds. For an overworked open-source maintainer, a report that says “your code has a problem” is an additional task. A report that says “here is the exact commit where this broke, here is working exploit code demonstrating the impact, and here is a suggested patch” is, as Anton Arapov of the OpenSSL Corporation put it in Anthropic’s announcement, “basically job done for an engineer.”
The Numbers Behind the Launch
Anthropic did not announce OSS Scanner today and claim it was ready. It announced it today after proving it works.
Over a period of six months, the program found more than 29,000 candidate vulnerabilities in widely used software. Staff have manually reviewed roughly 6,000 of those findings. Nearly 5,000 reports have already been sent to maintainers. The company is now replacing that manual review bottleneck with automated delivery — reports go directly to maintainers without a human intermediary, arriving faster but with a caveat that severity ratings may occasionally be wrong.
Anthropic says its expert penetration testers reviewed 97 critical and high-severity findings across 48 different projects and cleared 85 for immediate disclosure. Of the remaining 12, all but one turned out to be real bugs — they simply duplicated findings that had already been sent or were found by another researcher simultaneously.
wolfSSL, an early trial participant, reported that all but two of its 74 received reports were valid. Five of those reports became publicly registered CVEs.
Why Operational Technology Is the Hard Part
The Critical Infrastructure Defense Program is tackling a problem that has resisted standard cybersecurity approaches for years. Operational technology — the computers, controllers, and sensors embedded in industrial systems — is fundamentally different from the enterprise IT environments most security tooling is built for.
In enterprise IT, patching is a routine process. You push an update, reboot, done. In OT, the machine being patched might be controlling the pressure in a municipal water system or the voltage on a power distribution line. Taking it offline without disruption requires weeks of coordination. In some cases, the hardware is too old to receive modern patches at all. Booz Allen Hamilton’s Andrew Turner, president of commercial cyber, put it bluntly: operational technology is “the next frontier for autonomous AI-enabled attacks.”
That framing is significant. It acknowledges that attackers are already using AI to automate the discovery and exploitation of OT vulnerabilities — a point driven home by the ARTEX AI incident last week. Defenders, meanwhile, are still largely doing this work manually, with tools that were not built for the environments they are supposed to protect.
The CIDP is designed to close that gap by starting with the integrators and security vendors who already have the relationships and the site access that Anthropic lacks. Rather than selling Claude directly to utility operators — who face significant procurement, compliance, and operational risk hurdles — Anthropic is embedding its capabilities with the companies those operators already trust to touch their systems.
The Infrastructure Partners and What They Bring
The choice of eleven founding partners reflects the layered nature of OT security. Dragos and Nozomi Networks specialize almost exclusively in industrial control system security; they bring deep knowledge of OT protocols like Modbus, DNP3, and OPC UA. CrowdStrike and Palo Alto Networks bring endpoint and network telemetry at scale. Rockwell Automation and Hitachi build the actual industrial hardware, giving them insight into firmware-level vulnerabilities that no outside observer would normally see. The consulting firms — Accenture, Booz Allen, Deloitte, and PwC — bring the long-term operator relationships that determine whether any security tool actually gets deployed.
CrowdStrike’s Daniel Bernard noted that critical infrastructure facing “machine-speed threats” needs “machine-speed defense.” It is the kind of statement that sounds obvious but has been conspicuously absent from the AI industry’s self-description until very recently.
Open Source Gets Something Different
The OSS Scanner announcement is notable for what Anthropic is not asking for. There are no enterprise sales calls involved, no licensing discussions, no per-seat pricing. Projects apply by submitting a pull request to an Anthropic GitHub repository. Eligibility follows OSS-Fuzz’s existing standard: the project must have critical impact on infrastructure and user security, evaluated case by case.
Anthropic is funding the scanner’s operation through the Defender Advantage Fund (0xDAF), which it established in August. The company has also made direct investments in the Python Software Foundation, the Apache Software Foundation, and, through the Linux Foundation, Alpha-Omega and the Open Source Security Foundation.
Two coordination partners — Akrites and Gold Eagle — manage the vulnerability disclosure workflow, ensuring that findings are routed to maintainers through coordinated disclosure practices rather than dropped publicly before a fix exists.
What Comes Next
Anthropic is calling this launch a starting point. The CIDP is beginning with a small cohort of partners in order to learn which deployment strategies actually work in practice before expanding. The company says it will bring more partners and infrastructure sectors into the program over the coming months.
There are real open questions. Anthropic has not disclosed the commercial terms of the CIDP — it is unclear whether partners receive free model access or pay for compute. Axios raised a reasonable concern about how security teams will test and deploy AI-suggested fixes on running infrastructure without introducing new risk. These are not trivial problems.
But the broader context is important. Anthropic is the first frontier AI lab to build a dedicated, named program specifically focused on putting its models in the hands of infrastructure defenders rather than attackers. The timing — one week after an autonomous AI tool breached seven banks with no human involvement — is not coincidental.
The company believes AI will favor defenders within two years. Today’s announcement is the opening move in that argument.
The Anthropic Cyber Mission, including the Critical Infrastructure Defense Program and OSS Scanner, launched on October 8, 2026. Additional partners can register interest at anthropic.com/security.
