The most explosive news in the AI world in early April wasn't a new model release, but the sudden announcement that OpenAI, Anthropic, and Google—three companies usually competing for business—would join forces against Chinese AI firms.
Specifically, the three companies have begun sharing intelligence through the Frontier Model Forum (a group co-founded by these companies and Microsoft in 2023), targeting "distillation attacks"—where competitors use a large model's outputs to train their own smaller models.
Three Chinese companies were named: DeepSeek, Moonshot AI (Kimi's parent company), and MiniMax.
What is a Distillation Attack?
The method is straightforward: feed a target model a large number of carefully crafted prompts, collect the results, and use them to train your own model.
You don't need to know the target model's architecture or obtain its weights. Your model can learn a significant portion of the target's reasoning methods and knowledge structure. OpenAI suspected this when DeepSeek R1 was released; Sam Altman publicly stated at the time that DeepSeek had "improperly distilled GPT models."
Now Anthropic has provided specific numbers: these three Chinese companies interacted with Claude over 16 million times through approximately 24,000 fake accounts.
This is not normal user behavior. 16 million interactions represent industrial-scale, systematic data mining, carried out through a large number of fabricated accounts to evade platform terms of service detection.
Why Did Three Rivals Suddenly Become Allies?
Calling these three companies competitors is an understatement—OpenAI and Anthropic vie for the same enterprise contracts, while Google competes with both. However, when the scale of the problem exceeds what any single company's defenses can handle, competition temporarily takes a back seat.
The Frontier Model Forum's intelligence-sharing mechanism means that when one company discovers a new distillation attack pattern, the other two can synchronously update their defense strategies.
What did Anthropic contribute to this collaboration? It launched "Project Glasswing"—restricting access to its high-capability Claude Mythos model to specific partners, implementing stricter security detection mechanisms, and committing $100 million in usage credits to defense security partners like AWS, Google, and Microsoft.
What Exactly Did the Chinese Companies Steal?
According to Anthropic's allegations, the distillation covered multiple core capability areas, including computer vision, code generation, and complex reasoning.
A comparison of these companies' product roadmaps provides context:
| Company | Main Product | Link to Alleged Capability |
|---|---|---|
| DeepSeek | R-series reasoning models | Complex reasoning |
| Moonshot AI | Kimi long-context model | Document understanding, code generation |
| MiniMax | Video generation, multimodal | Visual understanding |
Each company's distillation targets may differ, but the common thread is gaining near-state-of-the-art capabilities without spending hundreds of millions of dollars on training.
What Does This Mean for Open-Source AI?
On a broader level, the formation of this alliance marks a shift in the AI industry toward an "intellectual property protection phase."
Previously, the focus was on AI openness, open-source, and model sharing. Now, Anthropic has directly restricted access to its most powerful model, and OpenAI is tightening enforcement of its terms of service.
If anti-distillation becomes an industry standard, researchers and small companies that rely on API calls to optimize their own models will also be affected—distinguishing "normal API calls" from "systematic distillation collection" is a technically and politically difficult problem.
As of now, none of the three named Chinese companies have publicly responded to these allegations.
Sources: CocoLoop, OpenAI, Anthropic, Google Unite to Combat Model Copying in China (Bloomberg); OpenAI, Google, Anthropic Unite Against AI Model Theft (Built In); OpenAI, Anthropic and Google Unite Against China AI Distillation (resultsense.com)