Anthropic Accuses Alibaba of 'Distillation Attack' on Claude AI Models

Anthropic Accuses Alibaba of ‘Distillation Attack’ on Claude AI Models Anthropic’s escalating clash with Alibaba is turning an obscure technical practice—“distillation attacks”—into a flashpoint in the global race over artificial intelligence and national security.

In late April, according to Anthropic, operators linked to Alibaba began systematically querying its Claude chatbot. Between April 22 and June 5, these actors allegedly carried out 28.8 million exchanges with Claude using almost 25,000 fraudulent accounts, a pattern Anthropic characterizes as a coordinated effort to “illicitly extract Claude’s capabilities” to train Alibaba’s own models.

As the activity unfolded, Anthropic’s policy team framed it not as routine competitive benchmarking, but as “the largest known distillation attack” on its systems to date. Distillation attacks, the company says, involve using a powerful model’s outputs to “harvest US AI capabilities … and repackage them as their own without incurring the training and R&D costs required to train US frontier models.”

The dispute moved into the political arena on June 10, when Anthropic’s head of policy, Sarah Heck, detailed the allegations in a letter to US Senators Tim Scott and Elizabeth Warren. She warned that such attacks could help Chinese models rapidly approach the capabilities of Claude’s advanced Mythos Preview system, which can detect software vulnerabilities and outperform humans on cybersecurity tasks.

Public reporting soon followed. The Financial Times highlighted Anthropic’s charge that Alibaba had obtained “illicit” access to Claude via fake accounts, while Business Insider emphasized that the company is “accusing China’s Alibaba of exploiting its AI models in a large-scale attack.”

Anthropic is now urging lawmakers to respond with new restrictions, including limits on China’s access to advanced US computing infrastructure and penalties for entities involved in such attacks. For Alibaba, already facing a US Pentagon blacklist and locked in a legal fight over that designation, the controversy adds further pressure as it tries to expand its Qwen large language models under Alibaba Cloud.

Continue reading https://foxvector.com/stories/019efe86-f805-1085-73b1-14e4298d2193