US Accuses Alibaba, DeepSeek of Systematic AI Model Theft

Model distillation and API-based knowledge transfer have become recurring friction points in open-weight AI development since 2023. Prior disputes typically involved private cease-and-desist communications or terms-of-service enforcement rather than coordinated government warnings. Historical precedents show that distinguishing between legitimate benchmarking, synthetic data generation for safety testing, and proprietary extraction is technically difficult when relying solely on output analysis. Previous industry conflicts over similar practices were generally resolved through platform-level access restrictions or licensing agreements rather than national security interventions. The pattern reflects a broader tension in the AI ecosystem where open-weight advancement relies on studying frontier model behaviors, yet proprietary developers view systematic output harvesting as an existential IP risk. Earlier cases lacked the synchronized timing between intelligence agency disclosures and private sector technical reports observed in this instance.

On September 9, 2026, United States security agencies issued a formal public warning accusing prominent Chinese artificial intelligence firms of systematically extracting proprietary knowledge from American technology companies. According to Bloomberg, the allegations specifically name Alibaba, DeepSeek, and Moonshot AI—the developer of the Kimi chatbot—as entities that have engaged in the systematic siphoning of US AI models. Security officials characterized these activities not as isolated incidents but as coordinated efforts to extract intellectual property, prompting an explicit warning to Silicon Valley developers to implement stronger protective measures for their assets. This government action represents a significant escalation in the scrutiny applied to cross-border AI development, moving beyond general trade concerns to specific accusations of model theft against named corporate actors.

The following day, on September 10, 2026, Anthropic released a detailed report that provided industry-level corroboration for the government's claims. According to TechCrunch, Anthropic’s report alleges persistent distillation campaigns conducted by Alibaba, Moonshot AI, and DeepSeek. The company stated that these attacks have escalated in recent months, correlating with intensified competition in the generative AI sector. Distillation, in this context, refers to the process where a smaller model is trained to replicate the outputs and behaviors of a larger, more capable proprietary model, effectively transferring knowledge without access to the original weights or training data. Anthropic’s documentation suggests that the accused firms utilized systematic querying strategies designed to map the decision boundaries and reasoning capabilities of frontier US models.

As of the current reporting window, none of the accused parties have issued public statements addressing the specific allegations made by US security agencies or the technical report released by Anthropic. Alibaba, despite being identified as a primary actor in both the government warning and the private sector report, has remained silent. Similarly, DeepSeek and Moonshot AI have not released comments denying or contextualizing the claims of systematic extraction. This lack of response creates an evidentiary gap regarding the intent and methodology of the accused firms. While the US government frames these activities as malicious theft, the technical reality of open-weight model development often involves learning from publicly available outputs, creating a zone of ambiguity between legitimate benchmarking and illicit IP extraction.

The sequence of events establishes a two-pronged pressure campaign: first, a national security framing by federal agencies, followed immediately by technical substantiation from a leading US AI lab. Bloomberg reports that the government's warning was directed at protecting American competitive advantages, implying that the alleged siphoning poses a direct threat to US technological leadership. The alignment between the government's public accusation and Anthropic's private sector findings suggests a coordinated information sharing environment, where intelligence agencies and frontier labs are collaborating to identify and publicize alleged adversarial machine learning activities. However, it remains unadjudicated whether the technical patterns identified by Anthropic constitute legal violations or merely aggressive competitive research within the current norms of AI development.

US frontier AI developers stand to benefit from strengthened IP protections and potential regulatory barriers against foreign competitors. Alibaba, DeepSeek, and Moonshot AI face immediate reputational damage and potential exclusion from US supply chains or API access. The magnitude includes three named companies representing significant portions of China's AI development capacity. If validated, these allegations could justify expanded export controls affecting billions in compute hardware and restrict access to foundational models. Conversely, if unsubstantiated, US firms risk credibility loss and retaliatory measures against American companies operating in China. The stakes extend beyond individual firms to the broader open-weight ecosystem, as precedent here may redefine acceptable research practices globally.