An artificial-intelligence model helping to build missiles sounds like science fiction. According to its own maker, it has already happened. In its threat intelligence report of 10 September 2026, the AI company Anthropic says a weapons cell in northern Yemen used its Claude models, including the coding tool Claude Code, to develop guidance and flight-control software for a guided rocket and two families of longer-range missiles.

The report is circulating again on social media as Houthi missiles strike Saudi Arabia. Here is what it says, what it does not say, and why it matters for anyone who follows missiles and air defence.

What the Cell Built

Anthropic says the cell ran three programmes. The first was a guided rocket using a commodity, phone-class flight computer with homing guidance in the final phase of flight. The second was a multi-stage ballistic missile with a stated range goal above 2,000 km. The third was a family of missiles referred to as the “R2000” set, including a hypersonic glide vehicle variant. We explain the differences between these weapon types in our guide to hypersonic weapons.

The work centred on guidance, navigation and control: the software that keeps a flying vehicle stable and steers it to a target. The report says the cell used Claude to fit an open-source autopilot onto the phone-class flight computer, write the control and position-estimation code, tune the control settings, run a firmware build pipeline and simulate flights. Other work included a six-degrees-of-freedom trajectory simulation for the ballistic missile and the use of reinforcement learning to tune flight control.

An AI Engineering Team

The striking detail is how the cell organised the work. According to the report, it ran several Claude instances at once and gave each a role, much as a lead engineer would delegate on a small team: one wrote the code, another did research and a third reviewed the first one’s output.

Anthropic says its safeguards blocked many of the cell’s requests, but not all. The operators hid their goals and what the software was for, and split their work across many sessions so that no single conversation revealed the whole programme.

A Test Firing That Failed

Anthropic says it has no evidence the cell succeeded in fielding an operational weapon. But it did test-fire a guided rocket. The test appears to have failed, and within hours the operators were back asking Claude to work out why. Anthropic found the activity through its own investigations into suspected weapons development, banned the accounts and shared information with government and industry partners.

That did not necessarily end the programme. The report says the cell had already built an offline simulation toolkit, compiled into a standalone program that runs without Claude or engineering software such as MATLAB.

Was It the Houthis?

Anthropic does not say. The report describes a cell “based in northern Yemen” and names no organisation. Northern Yemen, including the capital Sanaa, is controlled by the Iran-backed Houthis, which is why many reports have linked the two, and the Houthis do run long-range missile and drone programmes. But the link is an inference, not something the report states. Some social media posts also describe the cell as a “terrorist cell”; that is not Anthropic’s wording.

The timing explains the renewed attention. Since July 2026 Houthi missiles and drones have struck Saudi Arabia, including King Khalid International Airport in Riyadh, where a ballistic missile killed at least 13 people on 10 October. Casualty totals for the October attacks still differ between reports. We look at why the Houthis have survived a decade of air campaigns in Why the Houthis Are So Hard to Defeat From the Air.

Six Weapons Cases, Not One

Yemen is one of six conventional-weapons cases in the report, which Anthropic describes as a new category of misuse since its previous report in November 2025. The others are a Russia-based group building software for an autonomous FPV kamikaze drone swarm, tested in simulation and loaded onto real boards; a China-based actor who built a suite of about 16 modules for electronic warfare and suppression of enemy air defences; a China-based operation that drafted a fire-control specification for an anti-torpedo system; and two procurement and intelligence cases, one sourcing dual-use goods for Russian defence customers and one gathering public information on a directed-energy weapon.

Warum das wichtig ist

Guidance software is one of the harder parts of building an accurate missile, and it has traditionally needed skilled engineers. The report suggests that AI coding tools can stand in for some of that expertise for groups that already have hardware and know-how. Alongside the report, Anthropic’s Frontier Red Team published evaluations showing that models are making consistent progress on simulated weapons-development tasks, and the company says it has launched new classifiers to detect and block weapons-related traffic.

It also shows a new kind of intelligence source. Weapons programmes in Yemen have usually been reconstructed from missile debris and seized shipments, as in the reports of the UN Panel of Experts. Here, the developer of the tool the engineers used was able to watch part of the work in real time.

Diesen Beitrag auf Instagram ansehen

Sources: Anthropic, “Detecting and countering misuse of AI: September 2026” (10 September 2026), section on conventional weapons, case GTG-87001; Al Jazeera (11 September 2026); Associated Press; Benzinga (October 2026).