Anthropic accused DeepSeek and other Chinese AI labs of using Claude to train their models
Anthropic is accusing three Chinese AI labs, DeepSeek, MiniMax, and Moonshot AI (Kimi), of misusing its Claude model to improve their own AI products. The company claims the activity involved about 24,000 fraudulent accounts and more than 16 million interactions with Claude, which it describes as industrial scale campaigns. Anthropic says the goal was distillation, where a smaller model learns from a more capable model’s outputs, and argues that doing this without permission can give competitors a shortcut to advanced capabilities without comparable investment.
Anthropic warns that models produced through illicit distillation may not carry over key safety safeguards, increasing the risk of misuse. It says these systems could be applied to military, intelligence, or mass surveillance use cases, including offensive cyber operations or disinformation campaigns. Anthropic claims DeepSeek ran at least 150,000 targeted exchanges focused on Claude’s reasoning abilities and on generating censorship safe alternatives to sensitive topics, while MiniMax and Moonshot logged over 13 million and 3.4 million exchanges.
Anthropic is urging AI companies, cloud providers, and policymakers to address distillation, and suggests tighter limits on access to advanced chips to reduce the scale of these efforts. The claims land amid an ongoing US debate over chip export policies to China, and broader concerns about national security and economic risks tied to "unauthorized model extraction".
- AI Chatbot
- Freemium
- Proprietary
Related news
External links
- Detecting and preventing distillation attacksAnthropic • Official source
- Anthropic Accuses China’s DeepSeek Of Using Its Data To Train AI; Elon Musk Says Look Who’s TalkingNews18
- Anthropic Says DeepSeek, MiniMax Distilled AI Models for GainsBloomberg
- Anthropic accuses DeepSeek and other Chinese firms of using Claude to train their AIThe Verge
- Anthropic accuses Chinese AI labs of mining Claude as US debates AI chip exportsTechCrunch
Comments
The Chinese would never do such a thing. They're known for their integrity.
LOL
School: Students, keep studying very hard, and one day you'll be smarter than your professors.
Anthropic: Chatbots, keep studying very hard, and one day you'll be sued by us.
It’s rather amusing how large-scale web scraping — even when it includes copyrighted material — is framed as “transformative and legitimate learning” when done by a closed company, yet when open-source models distill the behavior of a proprietary system, it suddenly becomes “theft.” The underlying technique is the same: extracting statistical patterns from available data. Apparently, though, ethics shift depending on the software license. When the model is closed, learning is an exclusive privilege; but when open-source — that so-called “public commons” that fueled much of today’s innovation — comes close, intellectual property becomes a sacred cause. The irony is that the internet everyone trained on was treated as a commons, yet knowledge turns remarkably private the moment it escapes the walls.
100% agree and these Anthropic people are quite Arrogant and full of themseleves. First off you can go on hugging face.com and download from a huge selection off offline models that are uncensored without content filters and no lecture. Run them ooffline on your PC. Second off they seem to be a selfish company who is only concerned with lining the pockets of themselves to profit only themseleves and their people and Fuck everybody else! Well everything on the internet almost evrything is derived and brought to fruitin by the Open Source Community of Developers Worldwide and the tools these "selfish arrogant companys" access, these very same tools were not only built by the Open Source Community and maintained across varied platforms for consistent compatibility, without that they wouldnt have hardly anything to run their "claude" model on. The resources are not there without open source. It takes worldwide developers uniting together to build these massive projects and drivers and etc etc to make and host software that wil work out of the box for the layman pc owner and so on. And companys like anthropic who go "Oh look, lets micro manage A AI Aand lets charge people for its use and then also be closed source at the time." Well motherfucker i assure You that the rthings that are running in Claudes SOURCE CODE are OPEN SOURCE Code created by other developers for a variet of diff things yet you wanna use it in your code and hide behind a closed source wall you selfish Fu...! There is more in this world than to see how much you can prosper thourgh monetary gain. As your "novelty idea" to do what your doing with your AI Model will bring you low at some point and you will be filing bankruptcy. Deep seek has already set that in motion by showing that it doesnt take billions of dollars to create a AI from the ground up like these companys try and claim. The information and data that You are training your bots claude on is from the World Wide Web and the Data doesnt belong to you to begin with. So Fu...Anthropic right up their butt!