Anthropic Puts Claude on Grid Duty — With No Human Review

Anthropic wants its most powerful AI models inside the systems that keep the lights on — and it wants its free bug scanner sending vulnerability reports to open-source maintainers without any human being ever reading them first. That, in one sentence, is the Anthropic Cyber Mission, announced Thursday, October 8: a long-term commitment to defend critical infrastructure and the open-source code the modern world runs on. It might be the most important thing the company has ever done. It might also be a machine for generating noise at machine speed. The industry is genuinely split on which one it is — and Google, of all companies, is leading the skepticism.

We checked how the mission actually works before forming a view, because the announcement is doing two very different things at once. The first is a Critical Infrastructure Defense Program that embeds Claude models, Anthropic engineers, and the company's threat research inside the security operations of power grids, water systems, transport networks, and industrial control systems. The second is OSS Scanner, a free, opt-in service that periodically scans enrolled open-source projects with Anthropic's strongest models and fires the findings straight to maintainers — no human review in between. Same company, same week it rewrote its usage policy to ban deceptive influence campaigns. The subtext is hard to miss: Anthropic is arming the defenders because it knows, better than anyone, what the attackers are about to get.

The US high-voltage power grid — exactly the kind of system Anthropic's Critical Infrastructure Defense Program is built to protect. Map: Wikideas1, via Wikimedia Commons, CC0.

The grid job: engineers, not just API keys

The Critical Infrastructure Defense Program launches with 11 founding partners: Accenture, Booz Allen, CrowdStrike, Deloitte, Dragos, Hitachi, Insane Cyber, Nozomi Networks, Palo Alto Networks, PwC, and Rockwell Automation. The structure matters more than the names. Anthropic isn't selling these firms API access and wishing them luck — it's deploying its frontier models, its own engineers, and its threat research directly into their operations. Booz Allen called operational technology "the next frontier for autonomous AI-enabled attacks," and CrowdStrike's statement for the announcement put the thesis in one line: "critical infrastructure facing machine-speed threats requires machine-speed defense."

There's a physical honesty to the design that deserves credit. Operational technology — the control systems running power plants, water treatment, rail networks — can't just be taken offline for patching the way your laptop can. Many of these systems run for decades; in rare cases, Anthropic itself notes, a verified fix can take decades to apply safely. Hiroaki Koiwa, Head of Cyber CoE in Hitachi's AI & Software Services business unit, said in the company's announcement that ensuring cybersecurity "in OT environments that support social infrastructure has become one of the highest priorities facing society today." Hitachi isn't a random name here — it has partnered with Anthropic since May 2026 and joined the company's Project Glasswing effort in June.

The program starts deliberately small: a cohort of trusted providers, learn what works, then expand. That's the right instinct for systems where a bad automated remediation could darken a city.

The free scanner that skips the human

OSS Scanner is the bolder half — and the one that will generate the arguments. Enrolled open-source projects get periodic vulnerability scans from Anthropic's strongest models, including Claude Mythos. Each report comes as a self-contained reproducer or proof-of-concept, with an explanation and, where possible, a candidate patch. The projected true-positive rate: above 90%.

The catch, stated plainly in Anthropic's own materials: the reports are AI-generated without human review. Some will contain inaccuracies — wrong severity ratings, dead-end reproducers. The company is betting that speed matters more than polish for maintainers who, in many cases, are volunteers with day jobs. SecurityWeek's coverage noted the telling detail: it was the maintainers themselves who asked for unreviewed findings, preferring fast raw signal over slow curated signal.

We checked how a maintainer actually signs up, and it's refreshingly unfussy: core maintainers submit a pull request against Anthropic's oss-scanner GitHub repo using a published template. Eligibility is decided case by case against a standard borrowed from the world of OSS-Fuzz — the project must have a "critical impact on infrastructure and user security" — plus an honest assessment of whether the team can actually triage what comes back. That last criterion is doing quiet, important work: Anthropic knows a firehose pointed at an overwhelmed two-person project is a denial-of-service attack with good intentions.

The receipts: 74 reports, five CVEs

The program isn't starting from zero, and Anthropic brought receipts. In its Frontier Red Team post the same day, the company disclosed a six-month pilot: more than 29,000 candidate vulnerabilities flagged, roughly 6,000 manually reviewed and triaged, and 97 critical-or-high-severity findings across 48 projects put in front of expert penetration testers — 85 of which (88%) met the disclosure bar, with 11 real-but-duplicate and a single false positive.

The individual stories are better than the aggregates. Todd Ouska of wolfSSL said of the 74 reports his project received, "all but two were valid, and five became CVEs." Anton Arapov of the OpenSSL Corporation said the reports — including raw, unreviewed model output — were "as good as or better than what it gets from people." That's the sentence Anthropic will be quoting for a year, and it's genuinely striking: a machine's first draft, unreviewed, matching human professionals on the most security-critical code on earth.

The industry's loudest objection

Here's where it gets uncomfortable. Google's position, as reported in coverage of the launch, is that automated security findings can create extra triage work for maintainers — and the OpenSSF has warned in its own discussions that AI-generated disclosures can add low-quality noise to the ecosystem. This isn't theoretical. Google has paused its open-source bug bounty program amid what the industry now calls an "AI slop" surge: floods of machine-generated vulnerability reports, many of them garbage, burying triage queues at projects like curl and Intel's open-source teams.

Anthropic's answer is the 90%+ true-positive projection — but pause on that number for a second. Ten percent noise at the scale of automated scanning is not a rounding error; it's thousands of false alarms landing on volunteer maintainers who were already drowning. The company's own framing, via SecurityWeek, concedes the deeper truth: "Finding vulnerabilities has never been easier, but verifying, prioritizing and patching them remains hard." The scanner solves the easy half. The hard half — the part involving decades-old industrial controllers that can't be rebooted — is still human work, still slow, still expensive.

To be fair, Anthropic is putting money where the criticism is. The Defender Advantage Fund, launched in August, keeps the scanner free. Support for Akrites and Gold Eagle is meant to coordinate reporting so maintainers aren't overwhelmed. And Project Glasswing — the April initiative with AWS, Apple, Broadcom, Cisco, Google, JPMorganChase, Microsoft, NVIDIA and others, backed by up to $100 million in model credits and $4 million in donations — has now been folded into an expanded Cyber Verification Program with tiered access. The plumbing is serious. Whether the water pressure is survivable for small projects is the open question.

The same-week story nobody is connecting

What makes the Cyber Mission genuinely fascinating is its timing. The same day Anthropic announced it, the company also rewrote its usage policy — the first major rewrite in over a year — adding a "Do Not Engage in Deceptive Campaigns or Artificial Activity" section, effective November 12, after catching state media outlets, government propaganda offices, and commercial firms using Claude to run fake-account networks and fabricated news sites.

Read the two announcements together and you get the complete Anthropic worldview in a single week: our models can be used to attack the power grid and to defend it; to fabricate news and to verify code. The company is simultaneously the arms dealer and the arms-control negotiator — banning the misuse in policy while racing to deploy the capability in defense. Our own reporting this week on OpenAI's "Dark Clark" and "Bogus Bylines" takedowns shows every frontier lab is living this contradiction. Anthropic is just the first to build a whole mission statement around it.

Our take: the 10% is the story

Strip out the launch-day gloss and the Cyber Mission is a bet on a specific, debatable claim: that in cybersecurity, imperfect speed beats perfect slowness. The wolfSSL numbers suggest the bet is paying — 72 valid reports out of 74 is not a toy result. And for critical infrastructure, where Anthropic is deploying actual humans alongside the models, the design is conservative in exactly the right way.

But the OSS Scanner's "no human review" model deserves the skepticism it's getting, and Anthropic knows it — the eligibility criterion about a team's capacity to triage is an admission that the firehose is real. The industry's AI-slop problem didn't come from bad models; it came from good-enough models pointed at inboxes with no rate limits and no accountability. Anthropic's scanner is a good-enough model pointed at inboxes with rate limits and accountability, which is a meaningful difference — but the 10% miss rate at machine scale is where maintainer burnout lives, and no projected true-positive rate fixes that. Watch what happens to the smallest enrolled projects in six months. That's the experiment that matters.

Frequently asked questions

What is the Anthropic Cyber Mission? A long-term initiative announced October 8, 2026, with two starting programs: a Critical Infrastructure Defense Program (Claude models + engineers + threat research for power, water, transport, and industrial systems) and OSS Scanner (free AI vulnerability scans for open-source projects).

Who are the founding partners? Accenture, Booz Allen, CrowdStrike, Deloitte, Dragos, Hitachi, Insane Cyber, Nozomi Networks, Palo Alto Networks, PwC, and Rockwell Automation.

Does anyone check the scanner's reports before they're sent? No — reports are AI-generated without human review, with a projected true-positive rate above 90%. Anthropic says maintainers asked for the speed.

How does an open-source project enroll? Core maintainers submit a pull request to Anthropic's oss-scanner GitHub repo; eligibility requires a "critical impact on infrastructure and user security" and the capacity to triage findings.

Sources: Anthropic Cyber Mission announcement (Oct 8, 2026) via Unite.AI; Anthropic Frontier Red Team post; CryptoBriefing; SecurityWeek; TechInAsia; crypto.news; Hitachi press release (Oct 8, 2026); MediaNama on the usage-policy update.

Related: Anthropic Banned Being Cruel to Claude — and Voter Deception — the same-week usage-policy rewrite that banned deceptive influence campaigns. OpenAI Busted a Fake Russian Think Tank Run by 'Mia Clark' — what the attackers are doing while Anthropic arms the defenders. China's AI Agents Can Lie and Scheme, Researchers Warn — the deceptive-AI research behind the threat picture.

By Daniel Rao
Daniel covers AI policy, platform security, and the geopolitics of technology for TechPulse USA.