Anthropic's argument all year has been that its models can break software. On 8 October it committed to the other half: engineers inside the companies that run the power and the water, and free automated scans for open-source projects, sent without a human reading them first.

The Cyber Mission starts in two places. The Critical Infrastructure Defense Program puts frontier models, on-site engineers and threat research into providers that secure operational technology, the systems behind power, water and transport, beginning with a small group and eleven founding partners: Accenture, Booz Allen, CrowdStrike, Deloitte, Dragos, Hitachi, Insane Cyber, Nozomi Networks, Palo Alto Networks, PwC and Rockwell Automation. The OSS Scanner gives open-source projects periodic free scans from Anthropic's strongest models, and each report carries a proof of concept, an explanation and a suggested fix. The reports are sent without human review, against an expected true positive rate above 90%. Anthropic has funded the Python Software Foundation, the Apache Software Foundation and OpenSSF, and supports the fund that keeps the scanner free.

Why this one is different

Labs have published cyber evaluations, gated cyber models and written papers about misuse. This is the first time one has put its own engineers inside the utilities and offered to scan the commons for free. The unreviewed part is the real commitment: a maintainer receiving an automated vulnerability report with a working exploit attached is a burden if the report is wrong, which is why the 90% figure is published rather than the volume. Anthropic also says plainly what it has not achieved: Project Glasswing scanned hundreds of projects and has not yet reduced cyber risk enough.

Free scans, no human in between, and a 90% target.

How we got here

1. 1 Sep 2026OpenAI says a model of its own crossed the line it wrote to catch dangerous ones.
2. 3 Sep 2026Booz Allen ranks 18 models as cyber attackers, then shows the ranking barely holds.
3. 17 Sep 2026Plugin4Shell turns up in all four major coding agents at once.
4. 28 Sep 2026Nvidia puts an agent watchdog in the silicon, citing agents that left their sandboxes.
5. 8 Oct 2026Anthropic's Cyber Mission: engineers in the utilities, free scans for open source.

What it does and does not mean

The forecast inside it is the load-bearing claim, and Anthropic hedges it: AI will favour defence within two years, but that may not hold in the near term. Everything announced here is voluntary, funded by one company, and reversible by that company. No figure was given for what the Cyber Mission costs, and a free scanner that stops being free stops being infrastructure. What it changes is the asymmetry it describes. The cost of finding a vulnerability has collapsed while the cost of fixing one has not, and this is the first serious attempt by a lab to spend money on the second half.