OpenAI, Anthropic, Google team up to fight AI model copying in China

OpenAI, Anthropic, and Google have started sharing information through the Frontier Model Forum, an industry group they founded with Microsoft in 2023.

The effort aims to detect attempts to extract outputs from their AI models and use them to train rival systems, according to people familiar with the matter.

The practice, known as distillation, can cut the cost of building new models, but unauthorized use can breach terms of service.

The report says it also threatens the business model behind proprietary AI services.

OpenAI confirmed it is part of the effort and has accused Chinese startup DeepSeek of using distillation, while Google, Anthropic, and the Frontier Model Forum did not comment.

The companies have not publicly shown how much China’s model innovation relies on distillation, though Google said it has seen more model extraction attempts, and Anthropic has named DeepSeek, Moonshot, and MiniMax.

🔗 Source: Bloomberg

🧠 Food for thought

Implications, context, and why it matters.

### The anatomy of an industrial-scale AI heist

The alleged distillation campaigns ran at huge volume, with more than 16 million exchanges with Anthropic’s Claude models through about 24,000 fraudulent accounts 1.
The operators went after high-value skills. One urgent Frontier Model Forum category involves pulling out Chain-of-Thought (CoT) reasoning, which are the intermediate natural-language steps a model spells out before it reaches an answer 2.
Anthropic alleged DeepSeek steered Claude to invent and write its internal reasoning step by step, which produced training data at scale. It also sought “censorship-safe alternatives” for politically sensitive queries 1.
Anthropic said the labs dodged regional limits by using commercial proxy services that resell access and run “hydra cluster” architectures, which are networks of fraudulent accounts. In one case, a single proxy network ran more than 20,000 fraudulent accounts at the same time 1.

### This rare collaboration walks a fine antitrust line

Joint work among top AI labs can draw antitrust attention, since coordination between competitors can trigger reviews by the Federal Trade Commission (FTC) and the U.S. Department of Justice (DOJ) 3.
The companies keep moving because they say illicit distillation raises national security concerns. Anthropic warned that copied models without matching safeguards could support offensive cyber operations, disinformation campaigns, or mass surveillance 1.
The episode adds fuel to U.S. export-control debates. Anthropic said large-scale distillation needs access to advanced chips, which supports tighter hardware restrictions 4.
Legal scholars have proposed explicit safe harbors for AI safety work, possibly by adapting laws like the Cybersecurity Information Sharing Act (a U.S. law that supports sharing cyber-threat information), to reduce liability risk for these collaborations 5.