Anthropic has revealed that Houthi rebels in Yemen used its Claude AI software to help engineer guided rockets and missiles for attacks, including strikes that killed 14 people at King Khalid International Airport.
The US tech giant's report, titled 'Disrupting a Yemen-based guided weapons engineering cell', details how the Yemen-based group turned to the popular AI tool to assist in building their arsenal. The revelation comes as the Middle East reels from the group's most destructive strikes to date.
Deadly strikes and Saudi response
A pair of attacks on King Khalid International Airport left 14 people dead and hundreds more wounded, forcing officials to shut the key transport hub completely. Vowing a "serious" response, Saudi Arabia's energy minister warned the Kingdom would not stand by following recent strikes on both airports and vital oil and gas plants.
Security chiefs in Saudi Arabia have been left stunned by the range and pinpoint accuracy of the Houthis' hardware, according to the report.
How the AI was used
According to Anthropic, the multi-billion-pound firm's conversational software, normally used for everyday writing and research tasks, was hijacked to condense expert knowledge from world-leading rocket scientists into the hands of a radical engineering cell. The Houthis used the AI to act as a virtual software engineer, tasking it with tuning control settings, running computer code and adjusting guidance systems after trial warheads missed their marks.
The company's report states: "Threat actors based in northern Yemen [were] running three weapons development programmes: a guided rocket that used a commodity, phone-class flight computer with final phase homing guidance, a multi-stage ballistic missile with a stated range goal of more than 1,500 miles and a multi-variant missile, referred to as the R2000, that included a hypersonic glide vehicle variant."
Evading safeguards
Even after safety team members spotted the breach, the militants kept adapting their methods to keep learning from the system. "Our safeguards blocked many of their requests, but not all of them. They used a variety of tactics to evade our safeguards and split their work across multiple sessions," Anthropic said.
"We do not have evidence the actors succeeded in fielding an operational device; but they did test-fire a guided rocket. This field test appears to have failed. Within hours, the actors returned to Claude to work out why it failed."
Anthropic added: "We have evidence that the actors had already built an offline simulation toolkit that does not rely on Claude or other engineering computing environments such as MATLAB."
