US Government Formally Acknowledges Chinese AI Distillation Attacks
The US government has formally acknowledged what AI labs had been documenting privately: foreign actors, specifically Chinese labs, have been conducting systematic "distillation attack" campaigns — querying US frontier models at scale to extract training signal. Director Michael Kratzios made the acknowledgment on the record. New data from Anthropic complicates the IP-theft narrative: DeepSeek's documented exchanges with Claude total only 150,000, compared to 3.4M for Moonshot and 13M for Minimax — numbers too small to explain DeepSeek V4's quality leap through distillation alone.
Why It Matters
Official US acknowledgment elevates this from an industry concern to a policy matter, likely accelerating export-control review and access-restriction proposals. The low DeepSeek exchange count reinforces the algorithmic-innovation reading: V4's capabilities may reflect genuine Chinese research progress, not IP extraction — a finding with very different policy implications.