US Government Responds to Chinese Distillation Attacks on AI Models

The US government has announced measures to combat Chinese 'distillation attacks' aimed at stealing proprietary AI capabilities from American companies. Distillation attacks involve training less capable models using outputs from advanced models, which has been reported by OpenAI, Google, and Anthropic as a significant threat. In February, these companies disclosed that Chinese labs had executed millions of queries to extract information from their AI models. The White House memo outlines plans for private sector coordination, development of best practices, and diplomatic actions against China, which has denied the allegations. Despite these measures, experts express skepticism about their effectiveness in deterring further attacks. The memo notably lacks mention of strengthening semiconductor export restrictions, which have previously hindered Chinese AI development. The situation highlights ongoing tensions in the tech sector between the US and China.

Key Points: • The US is addressing distillation attacks by Chinese labs targeting AI models. • OpenAI, Google, and Anthropic reported millions of queries from Chinese actors. • The White House's response includes coordination and diplomatic actions, but lacks strong measures.