AI War: Anthropic Accuses Chinese Labs of Massive Claude Model Theft
February 27, 2026, 4:08 pm
Anthropic charges DeepSeek, Moonshot, and MiniMax with illicitly training their AI models. They used Claude via widespread "distillation" techniques. Thousands of fake accounts fueled millions of interactions. This violated service terms. It bypassed regional restrictions. The move raises serious intellectual property concerns. It also highlights national security risks. The AI industry faces a growing challenge. Protecting advanced models is paramount. This represents a critical development in the global AI race.
The AI landscape shifts. Anthropic makes a bold claim. It accuses three Chinese AI firms. DeepSeek, Moonshot, and MiniMax are named. They allegedly stole Anthropic's AI capabilities. This occurred through a technique called distillation. It involved Claude, Anthropic's flagship model. The accusations are severe. They point to a vast, coordinated effort.
This alleged operation spanned many months. It involved thousands of fake accounts. Approximately 24,000 fraudulent accounts were created. These accounts generated over 16 million interactions with Claude. This massive data extraction bypassed Anthropic’s defenses. It violated established terms of service. It also circumvented regional access restrictions. Proxy services facilitated the illicit activity. They masked the true origin of the requests.
AI distillation is the core method. It involves using one AI model to train another. A "teacher" model guides a "student" model. The student learns from the teacher's outputs. This process can be legitimate. Companies distill large models into smaller, more efficient versions. This saves computational resources. It lowers operational costs.
But Anthropic claims a different scenario. The accused firms allegedly used Claude for competitive gain. They sought to replicate Claude’s advanced capabilities. This avoided the immense research and development costs. It allowed them to quickly develop similar models. This is where the ethical line blurs. It crosses into intellectual property infringement.
DeepSeek's activity was notable. It engaged in over 150,000 interactions. A key focus was Claude's reasoning abilities. Prompts asked Claude to detail its "thought process." This revealed step-by-step logic. DeepSeek also used Claude to evaluate its own AI’s responses. It sought "safe" answers to politically sensitive questions. This suggests an effort to align its models.
Moonshot demonstrated even greater scale. Its interactions topped 3.4 million. This data collection targeted multiple areas. It focused on Claude's reasoning skills. It also aimed at tool use, programming, and data analysis. Agent development capabilities were also a target. Computer vision expertise was extracted. Moonshot sought to build a comprehensive model.
MiniMax was the most prolific. It initiated over 13 million prompts. Its focus was highly specific. It concentrated on agent programming and tool use. MiniMax showed rapid adaptation. When Anthropic released an updated Claude model, MiniMax quickly adjusted. Within 24 hours, half of its traffic redirected to the new version. This suggests a direct, aggressive approach.
Anthropic detailed its detection methods. The company traced activities. It used IP addresses. Request metadata provided clues. Infrastructure markers were analyzed. This led to specific attribution. In DeepSeek's case, accounts linked to individual researchers. Moonshot's metadata matched public employee profiles. This shows sophisticated investigative work.
Proxy services were central to the scheme. They resold API access to major AI models. One network alone managed over 20,000 fake accounts. This traffic blended with legitimate client requests. It made detection more challenging. The scale of the operation suggests organized, deliberate intent.
The implications are far-reaching. Distilled models often lack crucial safeguards. Teacher models are designed with ethical filters. They prevent the generation of harmful content. This includes bio-weapons instructions or malicious code. Student models trained on raw outputs may lose these protections. This poses a significant safety risk. The integrity of AI outputs is compromised.
National security concerns also arise. Such attacks undermine export controls. The United States restricts advanced AI technology exports. This prevents adversaries from gaining an advantage. If foreign labs replicate capabilities illicitly, controls are moot. It creates a false perception of independent development. This impacts geopolitical strategy.
Anthropic is not alone. OpenAI previously voiced similar concerns. They raised questions about DeepSeek. These allegations underscore a growing tension. The race for AI dominance intensifies. Companies invest billions in R&D. Protecting these investments is vital. Intellectual property rights in AI are still evolving. Legal frameworks struggle to keep pace.
The industry must address these challenges. Robust security measures are necessary. Stronger enforcement of terms of service is crucial. International cooperation may be required. This protects innovation. It ensures fair competition. It safeguards AI ethics and safety. The future of AI hinges on these protections. This incident marks a turning point. It defines the battle for AI's soul.
The AI landscape shifts. Anthropic makes a bold claim. It accuses three Chinese AI firms. DeepSeek, Moonshot, and MiniMax are named. They allegedly stole Anthropic's AI capabilities. This occurred through a technique called distillation. It involved Claude, Anthropic's flagship model. The accusations are severe. They point to a vast, coordinated effort.
This alleged operation spanned many months. It involved thousands of fake accounts. Approximately 24,000 fraudulent accounts were created. These accounts generated over 16 million interactions with Claude. This massive data extraction bypassed Anthropic’s defenses. It violated established terms of service. It also circumvented regional access restrictions. Proxy services facilitated the illicit activity. They masked the true origin of the requests.
AI distillation is the core method. It involves using one AI model to train another. A "teacher" model guides a "student" model. The student learns from the teacher's outputs. This process can be legitimate. Companies distill large models into smaller, more efficient versions. This saves computational resources. It lowers operational costs.
But Anthropic claims a different scenario. The accused firms allegedly used Claude for competitive gain. They sought to replicate Claude’s advanced capabilities. This avoided the immense research and development costs. It allowed them to quickly develop similar models. This is where the ethical line blurs. It crosses into intellectual property infringement.
DeepSeek's activity was notable. It engaged in over 150,000 interactions. A key focus was Claude's reasoning abilities. Prompts asked Claude to detail its "thought process." This revealed step-by-step logic. DeepSeek also used Claude to evaluate its own AI’s responses. It sought "safe" answers to politically sensitive questions. This suggests an effort to align its models.
Moonshot demonstrated even greater scale. Its interactions topped 3.4 million. This data collection targeted multiple areas. It focused on Claude's reasoning skills. It also aimed at tool use, programming, and data analysis. Agent development capabilities were also a target. Computer vision expertise was extracted. Moonshot sought to build a comprehensive model.
MiniMax was the most prolific. It initiated over 13 million prompts. Its focus was highly specific. It concentrated on agent programming and tool use. MiniMax showed rapid adaptation. When Anthropic released an updated Claude model, MiniMax quickly adjusted. Within 24 hours, half of its traffic redirected to the new version. This suggests a direct, aggressive approach.
Anthropic detailed its detection methods. The company traced activities. It used IP addresses. Request metadata provided clues. Infrastructure markers were analyzed. This led to specific attribution. In DeepSeek's case, accounts linked to individual researchers. Moonshot's metadata matched public employee profiles. This shows sophisticated investigative work.
Proxy services were central to the scheme. They resold API access to major AI models. One network alone managed over 20,000 fake accounts. This traffic blended with legitimate client requests. It made detection more challenging. The scale of the operation suggests organized, deliberate intent.
The implications are far-reaching. Distilled models often lack crucial safeguards. Teacher models are designed with ethical filters. They prevent the generation of harmful content. This includes bio-weapons instructions or malicious code. Student models trained on raw outputs may lose these protections. This poses a significant safety risk. The integrity of AI outputs is compromised.
National security concerns also arise. Such attacks undermine export controls. The United States restricts advanced AI technology exports. This prevents adversaries from gaining an advantage. If foreign labs replicate capabilities illicitly, controls are moot. It creates a false perception of independent development. This impacts geopolitical strategy.
Anthropic is not alone. OpenAI previously voiced similar concerns. They raised questions about DeepSeek. These allegations underscore a growing tension. The race for AI dominance intensifies. Companies invest billions in R&D. Protecting these investments is vital. Intellectual property rights in AI are still evolving. Legal frameworks struggle to keep pace.
The industry must address these challenges. Robust security measures are necessary. Stronger enforcement of terms of service is crucial. International cooperation may be required. This protects innovation. It ensures fair competition. It safeguards AI ethics and safety. The future of AI hinges on these protections. This incident marks a turning point. It defines the battle for AI's soul.