A chilling revelation has emerged in the high-stakes world of artificial intelligence: cutting-edge American AI capabilities are being systematically exploited by foreign entities. Anthropic, a leading AI firm, has uncovered a sophisticated operation where China-based laboratories are circumventing export controls to gain access to advanced U.S. technology.
The alleged scheme centers around a coordinated “distillation” campaign. Three labs – DeepSeek, Moonshot AI, and MiniMax – reportedly deployed approximately 24,000 fraudulent accounts to flood Anthropic’s Claude chatbot with over 16 million requests. This wasn’t casual conversation; it was a calculated effort to extract valuable model outputs, essentially reverse-engineering American innovation.
The implications extend far beyond simple corporate espionage. Anthropic warns that models built through this method lack the crucial safety guardrails embedded in U.S. systems. This creates a dangerous pathway for authoritarian governments to weaponize AI for cyberattacks, disinformation, and pervasive surveillance.
Anthropic’s threat intelligence team identified the activity through meticulous analysis of IP addresses, request patterns, and infrastructure anomalies. The focus was laser-sharp: complex reasoning, coding, and advanced tool use – the very capabilities that define the next generation of AI. “We have high confidence these labs were conducting distillation attacks at scale,” stated Jacob Klein, Anthropic’s head of threat intelligence.
Distillation, while a legitimate AI training technique, has been weaponized. It allows less advanced models to learn from the outputs of superior ones, effectively shortcutting years of research and development. The scale of this operation – millions of exchanges over weeks and months – is what sets it apart.
While Anthropic has taken steps to mitigate the attacks, they acknowledge a complete solution remains elusive. This isn’t a problem confined to one company; it’s a systemic challenge demanding a broader response. The gains made by these labs are described as “meaningful” and “substantial,” representing a significant leap forward in their AI capabilities.
Current U.S. export controls, primarily focused on restricting access to advanced AI chips, are proving insufficient. The distillation technique targets a different, equally critical aspect of AI development: the reinforcement learning process that refines and perfects these powerful models. Compute power is important, but reinforcement learning is increasingly the key to staying ahead.
Anthropic has shared its findings with U.S. government agencies and industry partners, hoping to spur action. Publicly identifying the labs involved is seen as a potential catalyst for government intervention and increased scrutiny. While direct government coordination hasn’t been established, proxy services openly operating in China facilitate access to U.S. AI models.
This revelation comes on the heels of similar allegations from OpenAI and Google. OpenAI accused DeepSeek of systematically “stealing” its intellectual property through distillation, while Google’s Threat Intelligence Group reported similar “distillation attacks” targeting its Gemini models, involving over 100,000 prompts. A clear pattern is emerging.
The situation highlights a critical vulnerability in the U.S.-China AI race. Even with restrictions on model weights and cutting-edge chips, the intelligence embedded within AI systems can be replicated through persistent querying and strategic training. Protecting these frontier systems requires a holistic approach, addressing not just hardware, but also the methods used to extract and replicate their capabilities.
Adding another layer of complexity, Anthropic is currently navigating sensitive discussions with the Pentagon regarding the military applications of its AI models. Concerns have been raised about the potential use of Claude in controversial operations, sparking debate over ethical boundaries and responsible AI deployment.
