Anthropic Accuses Chinese AI Firms of Industrial-Scale Model Distillation Attacks
Anthropic has uncovered a coordinated, industrial-scale effort by DeepSeek, Moonshot AI, and MiniMax to extract proprietary capabilities from its Claude models through fraudulent account usage and model distillation. The company says over 16 million interactions were generated in violation of its terms of service.
Anthropic Accuses Chinese AI Firms of Industrial-Scale Model Distillation Attacks
summarize3-Point Summary
- 1Anthropic has uncovered a coordinated, industrial-scale effort by DeepSeek, Moonshot AI, and MiniMax to extract proprietary capabilities from its Claude models through fraudulent account usage and model distillation. The company says over 16 million interactions were generated in violation of its terms of service.
- 2Anthropic Accuses Chinese AI Firms of Industrial-Scale Model Distillation Attacks On February 23, 2026, Anthropic, the AI safety-focused company behind the Claude family of large language models, publicly disclosed a sophisticated and large-scale campaign of model distillation allegedly conducted by three Chinese artificial intelligence laboratories: DeepSeek, Moonshot AI, and MiniMax.
- 3According to a detailed report published on Anthropic’s official news portal, these entities generated over 16 million API interactions with Claude through approximately 24,000 fraudulent user accounts, systematically harvesting model outputs to train and enhance their own proprietary AI systems.
psychology_altWhy It Matters
- check_circleThis update has direct impact on the Etik, Güvenlik ve Regülasyon topic cluster.
- check_circleThis topic remains relevant for short-term AI monitoring.
- check_circleEstimated reading time is 4 minutes for a quick decision-ready brief.
Anthropic Accuses Chinese AI Firms of Industrial-Scale Model Distillation Attacks
On February 23, 2026, Anthropic, the AI safety-focused company behind the Claude family of large language models, publicly disclosed a sophisticated and large-scale campaign of model distillation allegedly conducted by three Chinese artificial intelligence laboratories: DeepSeek, Moonshot AI, and MiniMax. According to a detailed report published on Anthropic’s official news portal, these entities generated over 16 million API interactions with Claude through approximately 24,000 fraudulent user accounts, systematically harvesting model outputs to train and enhance their own proprietary AI systems.
While model distillation—a technique in which a smaller, less resource-intensive model is trained on the outputs of a larger, more capable one—is a widely accepted and legitimate practice in the AI community, Anthropic asserts that the scale, intent, and method of these operations crossed into unethical and legally actionable territory. The company emphasizes that the attackers circumvented regional access restrictions, violated its Terms of Service, and engaged in deceptive account creation to bypass rate limits and monitoring systems designed to prevent abuse.
Anthropic’s internal security team detected anomalies in query patterns, including high-frequency, low-variance prompts designed to elicit specific types of responses—such as reasoning chains, code generation, and sensitive policy-guided answers—that are particularly valuable for distillation. These patterns were inconsistent with legitimate user behavior and pointed to automated, systematic extraction rather than organic usage. The firm’s investigation traced the traffic to IP ranges and infrastructure linked to the three accused firms, corroborated by digital forensics and behavioral analytics.
"Distillation is not inherently malicious," said a spokesperson for Anthropic in a statement. "But when conducted at industrial scale, using fraudulent means to bypass ethical and contractual safeguards, it becomes a form of intellectual property theft. We are not opposed to innovation—we are opposed to theft disguised as innovation."
Industry analysts note that this incident marks one of the most explicit and well-documented cases of cross-border AI model theft to date. "This isn’t just about copying outputs," said Dr. Elena Rodriguez, a senior AI ethics researcher at Stanford’s Center for Artificial Intelligence Policy. "It’s about reverse-engineering the decision-making architecture of a frontier model without access to its weights or training data. That’s a new frontier in AI espionage."
DeepSeek, Moonshot AI, and MiniMax have not issued public responses to Anthropic’s allegations as of press time. However, sources within China’s AI sector suggest that these firms have been under increasing pressure to close the performance gap with U.S.-based frontier models like GPT-4 and Claude 3. Some insiders indicate that distillation of Western models has become an open secret in certain Chinese AI labs, though few have been caught on such a large scale.
Anthropic has taken technical measures to mitigate further abuse, including enhanced bot detection, stricter geographic access controls, and dynamic watermarking of model responses to trace unauthorized usage. The company is also exploring legal avenues, including potential violations of the U.S. Defend Trade Secrets Act and international cybercrime conventions.
The incident raises broader questions about the governance of AI development in a globalized ecosystem. With no binding international treaties governing model training data or ethical distillation practices, companies are left to rely on terms of service and technical safeguards—both of which can be circumvented by determined actors. As AI models become more valuable than raw data, the lines between learning, imitation, and theft are increasingly blurred.
For developers and enterprises using Claude or other proprietary models, Anthropic urges vigilance. "We’re not just protecting our IP," the spokesperson added. "We’re protecting the integrity of the entire AI ecosystem. If unchecked, these practices erode trust, discourage investment in safety research, and ultimately harm users who rely on responsible AI."
As the global AI race intensifies, this case may serve as a watershed moment—pushing governments and industry consortia to establish clear norms around model training, data sourcing, and competitive fairness in artificial intelligence.
