Anthropic accuses Chinese AI firms of coordinated model‑distillation attacks

Anthropic released a detailed report on Thursday alleging that three China‑based artificial‑intelligence companies—Alibaba, Moonshot AI and DeepSeek—have been running systematic model‑distillation attacks against its proprietary systems. The company claims the campaigns have become more frequent and sophisticated as competition for cutting‑edge language models intensifies. TechCrunch reported that Anthropic described the activity as “persistent” and “escalating” over recent months.

Model distillation involves extracting the knowledge of a large, often proprietary, neural network and compressing it into a smaller model that can be reproduced or redistributed. While the technique can be used legitimately to create efficient versions of open‑source models, it also offers a shortcut for rivals to appropriate proprietary capabilities without incurring the research costs. When applied maliciously, distillation can effectively leak trade secrets and undermine a company’s competitive edge.

The accusations surface amid a broader backdrop of heightened tension between U.S. and Chinese AI developers. Over the past year, several high‑profile incidents—ranging from alleged data scraping to the unauthorized replication of model architectures—have underscored the difficulty of policing intellectual property in a field where code and datasets are easily shared across borders. Industry analysts note that the rapid commercialization of large language models has amplified incentives for aggressive tactics, especially as Chinese firms receive strong state backing to accelerate their AI ambitions.

In response, Anthropic said it is bolstering its defensive measures, including watermarking outputs and tightening access controls for its APIs. The company also announced plans to collaborate with external security researchers to develop detection tools that can identify distilled copies of its models in the wild. According to TechCrunch, Anthropic is exploring legal avenues to address what it describes as “unauthorized appropriation” of its technology.

The report has drawn attention from policymakers who are increasingly focused on AI security and cross‑border intellectual‑property enforcement. If the allegations are substantiated, they could prompt a wave of regulatory scrutiny aimed at curbing illicit model‑stealing practices and ensuring a more level playing field for AI innovators worldwide.

Stakeholders across the AI ecosystem are watching closely, as the outcome may shape how companies protect their most valuable assets and influence the future trajectory of the global AI arms race.