Anthropic Says Houthi Cell Tried Building Guided Missiles With Claude AI as Yemen Rebels Seize Red Sea Island

Since Iran-backed Houthi rebels intensified their offensive across Yemen this month, two separate developments have collided: an AI company's disclosure that a Yemeni weapons cell tried to use its chatbot to design guided missiles, and the rebels' capture of a strategic Red Sea island days later.

Anthropic said in a report published Thursday, September 10, that it had "identified a cell of threat actors based in northern Yemen running three weapons development programs" using its Claude AI models, according to The Hindu and Arab News. The programs included a guided rocket using a commodity phone-class flight computer with final-phase homing guidance, a multi-stage ballistic missile with a stated range goal above 2,000 kilometers (1,250 miles), and a multi-variant missile system that included a hypersonic glide vehicle.

Anthropic said the cell used its Claude Code tool to write guidance, navigation and control software that would otherwise have required a team of engineers. "We do not have evidence the actors succeeded in fielding an operational device; but they did test-fire a guided rocket," the company said. "This field test appears to have failed: within hours, the actors returned to Claude to work out why it failed." Anthropic banned the accounts once it detected the activity and did not name the group, though NPR and the Times of India both note it is likely tied to the Houthis, who control Sanaa and much of northern Yemen.

A broader threat report

The Yemen case was one piece of a larger Anthropic disclosure covering activity between December 2025 and August 2026, according to CBS News. The company said it also blocked three Iranian state-aligned accounts running influence operations that disguised AI-generated content as independent news across X, Instagram and TikTok. One Iran-linked account, CBS reported, was used to generate targeting recommendations against U.S. naval forces in the region and to build software for an Iranian domestic mass-surveillance system. Separate accounts tied to China and West Africa were used for surveillance and target identification, Anthropic said.

On the biological side, CNN reported Anthropic flagged roughly 35 "distinct research efforts" with potentially concerning activity over a 30-day window, including requests tied to gain-of-function research on the chikungunya virus, bird flu transmissibility in mammals, orthopoxviruses in the smallpox and mpox family, and novel venom toxins. Anthropic said those involved were "working scientists" and that it "cannot assert that they intended harm," since the same research could support vaccines or cures as easily as weapons.

A reasonable skeptic could look at 35 flagged research efforts, a failed but real missile test, and years of AI-safety promises and conclude voluntary company screening is not enough to stop determined bad actors, especially when Anthropic itself admits it cannot confirm intent in most cases. Anthropic's counter is that its safeguards worked exactly as designed: the accounts were caught, banned, and the rocket test failed. The report does not resolve which risk dominates going forward.

The ground war catches up

A day after the report surfaced, Houthi fighters captured the island of Mayun at the mouth of the Bab el-Mandeb Strait, confirmed by a senior official with Yemen's internationally recognized government and a Houthi official, both speaking anonymously to the Associated Press, as reported by NPR. About 12% of global shipping passes through Bab el-Mandeb, and it has served as Saudi Arabia's workaround since Iran began targeting vessels in the Strait of Hormuz.

Houthi military spokesman Yahya Saree said navigation "is safe for all companies except for Saudi vessels." United Nations spokesman Farhan Haq said Houthi forces also seized a compound in Mokha that had housed U.N. staff, who evacuated safely before the city fell. "No U.N. personnel remain in Mokha right now," Haq said.

Saudi Arabia shut down its East-West oil pipeline Friday as a "precautionary measure" after attacks the day before, according to the Saudi Energy Ministry, which did not identify who carried them out. That pipeline has carried exports that, per the International Energy Agency, more than doubled since the U.S. and Israel struck Iran on February 28. Separately, Iraq's government said Prime Minister Ali al-Zaidi ordered an investigation and dismissed the military commander overseeing operations in Maysan province after determining that attacks on Saudi Arabia were launched from a site there.

CNN and CBS covered Anthropic's report as a standalone AI-safety story, focused on model capability and company disclosure practices, without connecting it to the Houthi advance on the ground. NPR's reporting ties the AI disclosure directly to the live military situation in Yemen, giving readers the fuller picture of a rebel force that is simultaneously experimenting with AI-assisted weapons design and physically seizing chokepoints that squeeze Saudi Arabia's oil exports from two directions at once.

Anthropic has not said whether it referred the Yemen case to U.S. law enforcement or intelligence agencies, and no criminal charges or sanctions tied to the disclosed accounts have been announced. Whether the failed rocket test was a one-off or the group has since turned to other AI tools remains unknown, as does how the Mayun island seizure and pipeline shutdown will factor into Iran and Gulf states' scheduled Monday talks over Strait of Hormuz access.