American AI Giants Pool Data to Defend Against Attacks from Chinese Labs
OpenAI, Anthropic, and Google are officially sharing data for the first time through the Frontier Model Forum to counter Chinese labs. These companies are documenting extraction attacks via adversarial distillation—a technique where attackers send massive queries to donor models, collect responses, and train their own systems. The Forum, created in 2023 in partnership with Microsoft, has evolved from a platform for safety research into a channel for real-time sharing of violation intelligence.
Adversarial distillation differs from legitimate distillation, which is used to create compact model versions. The violation occurs when user agreement terms are ignored, allowing proprietary knowledge to be extracted without permission.
Scale of Attacks and Data on Chinese Players
In February 2026, the companies acted separately: OpenAI reported DeepSeek to Congress, Anthropic published an analysis of campaigns by DeepSeek, MiniMax, and Moonshot, and Google noted attempts to extract coding data and credentials from Gemini.
Anthropic detected over 16 million queries to Claude from 24,000 fake accounts via proxy services. Breakdown:
- MiniMax: over 13 million interactions;
- Moonshot: 3.4 million;
- DeepSeek: around 150,000.
Chinese labs bypass blocks since Claude is officially unavailable in China. These campaigns are coordinated and use commercial proxies for masking.
Blurred Lines and Criticism of the Approach
Experts note the ambiguity of the accusations. Professor Eric Cambria from Nanyang Technological University highlights the blur between legitimate API use and exploitation. Lia Raquel Neves from AI ethics consulting views the complaints as ToS violations, not industrial abuses.
Critics point to hypocrisy: Anthropic itself used distillation and public data for its models. American companies distill proprietary systems into client versions themselves, blurring the moral line.
Political Context and Recommendations
The issue has reached the government level. The IAPS Institute in March proposed adding the implicated labs to the US Department of Commerce's Entity List, imposing sanctions under the PAIP Act, and developing an NIST framework for distillation protection. The first PAIP Act sanctions were applied on February 24, 2026.
CSET analysts identified over 9,000 PLA tenders in 2023–2024 requesting AI applications, including DeepSeek models. This ramps up pressure to escalate technical complaints into regulatory measures.
Key Takeaways
- OpenAI, Anthropic, and Google are coordinating data sharing through the Frontier Model Forum to track adversarial distillation from DeepSeek, MiniMax, and Moonshot.
- Chinese labs generated >16 million queries to Claude from 24,000 fake accounts via proxies.
- Attacks violate ToS, but the line with legitimate distillation is blurred; critics see selectivity.
- Political escalation: proposals for PAIP Act sanctions and Entity List based on PLA military tenders.
- Shift from individual complaints to collective monitoring underscores the threat to proprietary models.
This collaboration sets a precedent for the industry, where protecting intellectual property in AI becomes a collective task. Middle/senior-level developers should factor in API leak risks when designing security systems.
— Editorial Team
No comments yet.