Anthropic's Cyber Mission puts Claude models and engineers inside critical infrastructure defence
Anthropic launches a two-track cyber programme: a Critical Infrastructure Defense Program with 11 partners and a free OSS vulnerability scanner.
On 8 October 2026, Anthropic announced the Cyber Mission, a structured, long-term commitment to helping defenders secure the systems that modern life depends on. It covers two specific areas: critical infrastructure such as power grids, water systems, and transport networks, and the open-source software that underpins almost every application in existence. Both tracks reflect lessons Anthropic says it learned from Project Glasswing, an earlier pilot that ran from April 2026 and has now been folded into a larger programme.
What happened with Project Glasswing?
Glasswing was Anthropic’s first serious attempt to put frontier AI models to work for cyber defenders rather than attackers. Between April and July 2026, partners working through the programme verified at least 129,000 software vulnerabilities. Anthropic’s own open-source scanning added around 5,500 more between April and October.
That sounds impressive, but it also exposed a bottleneck. Of more than 29,000 candidate vulnerabilities identified over six months of scanning, only around 6,000 received manual review and triage. Finding vulnerabilities at scale turns out to be the easier part. Verifying, prioritising, and actually fixing them is where the real resource crunch sits. The Cyber Mission is Anthropic’s answer to that problem.
Track one: the Critical Infrastructure Defense Program
The Critical Infrastructure Defense Program (CIDP) brings together 11 founding partners: Accenture, Booz Allen, CrowdStrike, Deloitte, Dragos, Hitachi, Insane Cyber, Nozomi Networks, Palo Alto Networks, PwC, and Rockwell Automation.
The proposition is not just API access to Claude. Anthropic is committing frontier models, on-site engineers, and threat research to a small initial cohort of organisations defending operational technology (OT) environments. Power grids, water treatment facilities, and transport networks often run on decades-old proprietary equipment where an incorrect software change can have physical consequences. That context shapes everything about how the programme works.
Security vendors and service providers that build products or services for critical infrastructure can register their interest in the CIDP as it expands beyond the founding cohort.
Anthropic also notes that its earlier government cyber defence work has already reached more than half of US states since June 2026, with Claude being used in code scanning, patch development, incident response, and red teaming workflows.
Track two: OSS Scanner
The second track is aimed at open-source maintainers. Almost all software depends on open-source code, much of it maintained by small volunteer teams working without dedicated security resources. Anthropic’s free, opt-in OSS Scanner periodically scans enrolled projects using its strongest models and sends back a report containing an explanation of the flaw, a proof-of-concept exploit, and a proposed patch where one is available.
The service is inspired by Google’s OSS-Fuzz, which has run automated fuzzing against open-source projects for years. Anthropic says it applies additional token-intensive experimental harnesses to dig for deeper bugs, and it expects a true-positive rate above 90%.
There is one important caveat to understand: these reports are sent without human review before delivery. Anthropic is transparent about this. Some reports will contain inaccuracies, including wrong severity ratings. There is also no 90-day disclosure deadline applied to unvalidated findings. Maintainers who enrol are taking on the responsibility of checking each report’s accuracy and assessing its impact before acting on it. For a busy volunteer maintainer, that is a real consideration, but the alternative, receiving no automated scanning at all, is worse for most projects.
Core maintainers of critical open-source projects can enrol at red.anthropic.com/oss-scanner.
What is funding this?
The Defender Advantage Fund (0xDAF), which Anthropic launched in August 2026, is what keeps OSS Scanner free and supports pilot programmes across both tracks. The fund provides $35 million in Claude credits, with grants focused on patching live vulnerabilities, automating scanning and patching workflows, and helping projects pursue structural security improvements that reduce exposure to whole classes of attack rather than individual bugs.
Anthropic has already directed funding through 0xDAF to organisations including the Python Software Foundation, the Apache Software Foundation, Alpha-Omega, and OpenSSF through the Linux Foundation.
The expanded Cyber Verification Program
Alongside the Cyber Mission, Anthropic rebuilt its Cyber Verification Program (CVP) on 6 October 2026 into three tiered access levels. Project Glasswing members are transitioning to the top tier, called Specialised Access, which is reserved for verified organisations authorised to test systems that could affect people’s lives or markets, including power grids, flight operating systems, and interbank transfer infrastructure. All tiers include access to Claude Opus 5.5, Claude Sonnet 5.5, and Claude Mythos 5.1, with reduced safety restrictions proportional to the level of vetting completed.
What does this mean for you?
If you work in critical infrastructure security or sell products into that sector, the CIDP is the most direct route to accessing frontier model capabilities with Anthropic engineers alongside you. The programme is deliberately starting small, so early registration matters if you want to be in the first wave.
If you maintain or contribute to open-source software, the OSS Scanner is genuinely useful, free, and asks nothing more than an opt-in. The lack of human review before reports arrive is a real consideration, but a 90-plus percent true-positive rate from a model with access to experimental vulnerability harnesses is a meaningful addition to any project’s security posture. Check what arrives, do not treat it as gospel, and patch what is real.
For the broader enterprise security community, the programme signals where Anthropic is concentrating effort: not on selling AI to attackers or building generic security features, but on giving defenders with specific, verified mandates access to the most capable models with fewer restrictions. That focus is deliberate, and the disruption of an AI-assisted Chinese state-sponsored espionage campaign that Anthropic disclosed separately in October 2026 makes clear why the company thinks the stakes justify the approach.