Anthropic has accused Chinese tech giant Alibaba of illicitly extracting capabilities from its Claude AI model, according to a Reuters report published June 24, 2026. The San Francisco-based AI safety company says the extraction was carried out without authorization, raising immediate questions about how large language model intellectual property can be protected in an era of rapidly proliferating AI tools.
The non-obvious detail buried in this story: the alleged method wasn’t a traditional cyberattack or data breach. Anthropic’s claim centers on a technique known as model extraction — where an adversary queries a model repeatedly and systematically to reconstruct its behavior, effectively cloning its capabilities without ever accessing the underlying weights or source code directly.
How Anthropic Says Alibaba Extracted Claude AI Capabilities
Model extraction exploits the public-facing API of a deployed AI system. By sending thousands — sometimes millions — of carefully crafted prompts and logging the outputs, a sophisticated actor can train a separate model to mimic the original’s responses. The resulting system isn’t a perfect copy, but it can replicate enough of the original’s reasoning style, knowledge, and instruction-following behavior to be commercially useful.
Anthropic alleges that Alibaba used this approach to siphon Claude AI capabilities and funnel them into competing products. Alibaba has not publicly confirmed the accusation. Reuters reported that Anthropic’s complaint represents one of the most direct and detailed allegations of AI model theft made by a major Western AI lab against a Chinese technology firm to date.
Anthropic declined to specify exactly which Claude model version was targeted or over what time period the extraction allegedly occurred. That ambiguity is itself telling: model extraction attacks are notoriously difficult to detect because individual API queries look indistinguishable from normal developer or researcher traffic.
Why This Goes Beyond a Corporate IP Dispute
Anthropic was founded in 2021 by former OpenAI researchers with an explicit mission around AI safety. The company has attracted billions in investment — including from Google and Amazon — and positions Claude as a safer, more interpretable alternative to competing large language models. An adversarial extraction of Claude’s capabilities doesn’t just hurt Anthropic commercially; it potentially transfers the alignment and reasoning properties that Anthropic spent years developing into a system with no safety audit trail.
The broader AI industry has long flagged model extraction as an underregulated threat. Unlike source code, which is protected by copyright the moment it’s written, the legal status of a model’s learned behavior — its “capabilities” — sits in murky territory. Courts in the U.S. have not yet definitively ruled on whether extracted model behavior constitutes theft of intellectual property, trade secrets, or something else entirely.
This dispute lands against a backdrop of intensifying U.S.-China competition in AI. As China has moved aggressively to project technological influence abroad, American AI labs have grown increasingly wary of state-adjacent actors accessing their systems. Alibaba’s cloud division, Alibaba Cloud, operates internationally and offers API access to a range of AI services, giving it legitimate reasons to interact with third-party AI systems — which is precisely what makes policing this kind of activity so difficult.
Anthropic’s Legal and Technical Options
Anthropic has several potential avenues. On the legal side, a trade secret misappropriation claim under the Defend Trade Secrets Act is the most likely vehicle if the company pursues U.S. litigation. To succeed, Anthropic would need to show that the extracted information constitutes a protectable trade secret and that Alibaba’s methods constituted improper means — a bar that model extraction cases have not yet been forced to clear in federal court.
On the technical side, AI labs have developed rate-limiting, query watermarking, and output perturbation tools specifically to make model extraction harder. Watermarking, in particular, has matured rapidly: researchers can now embed statistical signatures into a model’s outputs that survive the extraction process and appear in the cloned model’s responses — essentially a fingerprint that proves lineage. Whether Anthropic had such protections in place on the affected Claude version has not been disclosed.
The geopolitical dimension adds another layer. U.S. export controls have already restricted the sale of advanced AI chips to China, but they don’t cover behavioral extraction from publicly accessible APIs. Congress has been briefed on model extraction risks before, though no legislation specifically targeting the practice has passed.
What Alibaba Has — and Hasn’t — Said
As of publication, Alibaba has not issued a detailed rebuttal to Anthropic’s allegations. The company’s AI division has been aggressively expanding its own large language model lineup, including the Qwen series, which has drawn attention from researchers for its competitive performance on international benchmarks. Whether any of that performance derives from extracted Claude AI capabilities is precisely what Anthropic’s complaint puts in question.
The absence of a denial isn’t evidence of guilt — Alibaba may be preparing a formal legal response. But the silence has kept Anthropic’s framing dominant in the initial news cycle, and the story has already prompted calls from AI researchers on social platforms for clearer international norms around model intellectual property.
Anthropic’s next move will determine how far this goes. If the company files a formal lawsuit in a U.S. court, it will force a legal reckoning with model extraction that the AI industry has deferred for years. A settlement or quiet resolution, by contrast, would leave the legal landscape exactly as undefined as it is today — and signal to other actors that the cost of extraction remains low.
