Intellectual Property Theft

The US and China are in an AI supremacy race. Accusations of “distillation,” a technique where less advanced AI models learn from superior ones, have surfaced, with the US claiming China uses it to steal intellectual property and bypass R&D costs. However, some experts, like Cohere CEO Aidan Gomez, argue this overlooks China’s independent innovation and world-class capabilities, suggesting the US lead is rapidly diminishing. China refutes these claims, citing its own advancements and resilience despite chip restrictions.

Anthropic’s head of threat intelligence, Jacob Klein, reports that Chinese companies are increasingly using illicit distillation to steal AI model technology. This practice involves exploiting proprietary models like Anthropic’s Claude to train competing technologies at lower costs. Klein cites Moonshot AI’s Kimi K3 model as an example of technology illicitly trained using Claude. The issue raises concerns about intellectual property theft and national security risks.

Anthropic accuses Alibaba of a massive “illicit distillation attack” to steal its AI capabilities, using millions of fraudulent interactions. This follows recent White House directives and Anthropic’s previous alerts about similar theft attempts. The accusation highlights growing concerns over AI intellectual property theft and national security in the global AI race.

Anthropic has identified a large-scale operation extracting proprietary capabilities from its AI model, Claude, through deceptive accounts and sophisticated evasion tactics. This “distillation” campaign, conducted by overseas laboratories, aims to rapidly acquire advanced AI functionalities. The illicitly trained models pose national security risks by bypassing safety guardrails. Anthropic advocates for multi-layered defenses, including behavioral fingerprinting and traffic classifiers, to combat these extraction efforts and calls for cross-industry collaboration.