Anthropic accuses Chinese rival Alibaba of illicitly extracting AI capabilities

US artificial intelligence (AI) giant Anthropic has accused Chinese e-commerce and technology firm Alibaba of “brazenly” and “illicitly” extracting its Claude AI model’s capabilities.

In a letter sent to two members of the US Congress, the San Francisco-based company said operators linked to Alibaba carried out almost 29 million exchanges with Claude using thousands of fraudulent accounts in what it called the largest extraction campaign of its kind.

Anthropic urged Congress to penalise the companies behind attacks like this and to ramp up measures to prevent US tech from being stolen.

The BBC has contacted Alibaba for comment and requested more details from Anthropic.

Anthropic’s letter, dated 10 June and addressed to US Senators Tim Scott and Elizabeth Warren, accused New York Stock Exchange-listed Alibaba of carrying out “the largest campaign to illicitly extract Claude’s capabilities”.

According to Anthropic, the campaign was carried out through what are known as “distillation attacks”, which extracted answers from a stronger AI model to train a weaker one.

Alibaba-linked operators targeted Claude’s most valuable capabilities, including its ability to tackle longer and more complex tasks and its approach to decision-making, Anthropic said.

These type of attacks are carried out on an “industrial scale” to enable Chinese companies to harvest and repackage US AI capabilities as their own, the company said.

The letter also cited other alleged attacks, which Anthropic said posed a threat to the US military.

“Distillation attacks turn hundreds of billions of dollars in American investment and [research and development] into a massive subsidy for our geopolitical competitors,” said Anthropic.

It cited the US Department of Defense’s claims that Alibaba and several major firms like car maker BYD and tech company Baidu are tied to the Chinese military.

The companies have denied any such allegations, while Alibaba this week sued the US government in a bid to get its name removed from the Pentagon blacklist.

US developers have previously accused Chinese competitors of using distillation attacks to train their models to rival American AI technology at a fraction of the cost.

OpenAI has also previously accused Chinese groups of employing the same practice.

Anthropic is a leading AI developer and, alongside ChatGPT-maker OpenAI, is gearing up for a blockbuster stock market debut that could make it one of the most vaulable companies in the world.

But some of Anthropic’s more advanced models, such as Mythos, have raised cybersecurity concerns over their ability to target weaknesses in computer systems.