Could AI really help build a missile? Anthropic says a group in Yemen tried to do exactly that.
It used its Claude chatbot to develop software for guided rockets and missiles.
In a report, the US AI company said it uncovered a threat cell in northern Yemen working on three weapons programmes.
These included a guided rocket, a ballistic missile designed for a range of more than 2,000km, and a missile involving a hypersonic glide vehicle.
Claude was reportedly used to develop guidance and navigation software that would normally require a team of engineers.
Anthropic said there was no evidence the weapons became operational, although the group did test-fire a guided rocket.
AI Misuse Raises Wider Concerns
The test apparently failed. Within hours, the developers reportedly returned to Claude to figure out what went wrong.
Anthropic said it banned the accounts once it detected the illegal activity.
While it did not name the group, the report points toward the Houthis, who control much of northern Yemen.
The disclosure comes as Houthi forces expand their offensive around the strategic Bab al-Mandab shipping route.
Anthropic also reported attempts involving biological weapons and cyber espionage.
The bigger question is hard to ignore: as AI becomes more capable, how do we make sure its power stays in the right hands?
