A new Anthropic report alleges Alibaba, Moonshot AI, and DeepSeek have run persistent distillation campaigns against its models as competition intensifies.
Anthropic published findings Thursday alleging that China-based AI companies have escalated efforts to extract training signal from its models through systematic distillation attacks. The report frames this as part of a broader pattern of IP extraction as the gap between US and Chinese frontier models narrows.
Distillation, where a smaller model learns by querying a larger one's outputs, sits in a legal and ethical gray zone that labs are now trying to police through terms of service and technical countermeasures rather than courts.
This is the clearest public signal yet that frontier labs view model distillation as a competitive and security threat, not just an academic curiosity. Expect tighter API rate limits, watermarking, and access tiers aimed at foreign competitors, plus pressure for export-control-style policy on model outputs.
The daily signal, curated. Get it in your inbox.
Subscribe on LinkedIn →