Yemen missile cell used AI to write guidance software, Anthropic says

Yemen missile cell used AI to write guidance software, Anthropic says / bne IntelliNewsFacebook

TwitterLinkedInTelegramFlipboardFeedly

By bnm Tehran bureau September 11, 2026

A weapons cell in northern Yemen used Claude to write guidance software for a guided rocket and a 2,000 km ballistic missile, Anthropic said on September 10.

The US AI company's threat report does not name the Houthi movement, which controls northern Yemen and the missile arsenal there. But the programmes described match the group's public claims: a multi-stage ballistic missile, a hypersonic glide vehicle variant and a cheap tactical rocket with terminal homing. The finding matters because a model provider, rather than a UN panel working from recovered wreckage, has identified an active missile programme from the inside, and because the cell's work continued offline after its accounts were banned.

In place of software engineers, the cell used Claude Code to write the guidance, navigation and control software that steers a flying vehicle. It integrated an open-source autopilot onto a commodity phone-class flight computer, wrote the control and position estimation code, ran a firmware build pipeline and performed flight simulations. The actors managed several instances of the model at once, assigning one to write code, another to research and a third to review the code. A third programme, which the cell called the "R2000" set, covered several missile variants.

The cell test-fired a guided rocket in Yemen. The test appears to have failed, and within hours the actors returned to the model with telemetry to work out why. Anthropic said it has no evidence the cell fielded an operational device. It found the actors had already compiled an offline simulation toolkit that runs without Claude, and had used reinforcement learning to tune flight control and calibrate a digital model of an operational weapons system to reduce dependence on physical testing.

Safeguards blocked many requests. The actors evaded others by hiding the end product and splitting work across sessions so no single conversation revealed the full intent. Anthropic banned the accounts and passed threat information to public and private sector partners. The Yemen case is one of six conventional weapons cases in the report, alongside three in China and two in Russia.

The report also details three Iranian state-aligned influence operations, attributed to the Islamic Culture and Communications Organization under the Ministry of Culture and Islamic Guidance, the Islamic Propaganda Office of Khorasan Razavi, which ran a "cognitive warfare" command room from a Mashhad seminary, and the Bina Cultural Observatory of the Islamic Propaganda Organization.

The operators tied their work to the state doctrine of "Jihad al-Tabyin", or explanatory jihad. Deliverables included plans for the funeral of Iran's Supreme Leader and messaging in the voice of an IRGC spokesman during the 2026 US-Israel-Iran war that attributed false claims to CSIS, Brookings and RAND. Access to Claude is blocked inside Iran, so the operators used VPNs and foreign phone numbers, then named their institutions in conversation.

A separate operation linked to the exiled People's Mojahedin Organisation of Iran cloned a real activist's Telegram account, read 8,400 of his posts to copy his style and held live political conversations with his contacts inside Iran. At least four people running it work for official NCRI media outlets.

A further campaign is attributed with high confidence to UAE government officials. A single operator ran an AI persona called "Deadshot" under a doctrine file describing "a coordinated transatlantic and regional operation to dismantle the Muslim Brotherhood globally", managed about 300 fake influencer accounts, cloned a Swiss NGO's identity, ghost-wrote testimony for the 62nd session of the UN Human Rights Council and profiled 18 MEPs and UN rapporteurs critical of the UAE's conduct in Sudan.

Anthropic also banned an account in June that was building a platform to profile social media users in Iran and the Gulf, linked to S2T Unlocking Cyberspace, an Israeli-Singaporean intelligence vendor. The platform produced briefings in formal Arabic styled as government reports and held a stock of more than 255 synthetic accounts.