June 25, 2026, (Inside AI) — Anthropic, the U.S. artificial intelligence company, has formally accused Chinese technology conglomerate Alibaba of unlawfully extracting capabilities from its Claude AI model. The allegation, detailed in a letter reviewed by Reuters, describes the incident as the largest known attack of its kind against the firm.
The letter asserts that Alibaba engaged in systematic extraction of Claude's proprietary functionalities. Anthropic claims this was done without authorization, violating its terms of service and intellectual property rights. The exact methods alleged remain undisclosed, but such attacks often involve querying a model at scale to replicate its behavior.
This accusation lands amid escalating U.S.-China tensions over AI dominance. Both nations view advanced AI as critical to economic and military superiority. The incident could intensify scrutiny of cross-border AI interactions and data security.
Anthropic, founded by former OpenAI researchers, positions itself as a safety-focused AI developer. Its Claude models compete directly with offerings from OpenAI and Google. Alibaba, meanwhile, has invested heavily in its own large language models, including Tongyi Qianwen, aiming to rival Western counterparts.
The alleged extraction likely targeted Claude's ability to handle complex reasoning, nuanced conversation, or safety guardrails. Model extraction attacks typically involve sending numerous prompts to map out a model's responses, then training a new model to mimic those outputs. This can undermine competitive advantage and potentially expose safety mechanisms.
Neither company has publicly released the full letter. Anthropic declined to comment beyond the document. Alibaba did not immediately respond to requests for comment. Reuters reporter Syakir Jasnee broke the story, citing the letter as the primary source.
If proven, the incident would mark a significant escalation in AI intellectual property disputes. It follows a pattern of alleged technology transfer between U.S. and Chinese firms, often entangled in geopolitical friction. The scale claimed by Anthropic suggests a coordinated effort rather than casual probing.
Industry observers note that model extraction is notoriously difficult to prove. Distinguishing between legitimate use and illicit copying requires forensic analysis of query logs and output patterns. Anthropic's confidence in labeling this the "largest known attack" hints at substantial evidence.
Legal frameworks for AI IP protection remain nascent. U.S. law offers some recourse under the Computer Fraud and Abuse Act or trade secret statutes, but applying them to model extraction is untested. International enforcement adds complexity, especially given strained U.S.-China relations.
The accusation could ripple through the AI industry, prompting companies to tighten API access and monitoring. It may also fuel calls for clearer regulations on AI model protection. For now, the letter serves as a stark warning about the lengths competitors might go to close the AI gap.
