San Francisco and London, 26 March 2025. In a dramatic escalation of the global AI arms race, Anthropic, the frontier AI lab behind the Claude family of models, has formally accused Chinese e-commerce titan Alibaba of systematically extracting proprietary intellectual property from its systems. The accusation, filed in a strongly worded legal submission to the Northern District of California, alleges that Alibaba’s cloud unit used sophisticated network probing and model interception techniques to reverse-engineer parts of Anthropic’s safety stack.
Anthropic’s complaint, obtained by this correspondent, details a series of unauthorised intrusions beginning in late 2024. The attackers allegedly exploited weak links in third-party API aggregators to send carefully crafted queries designed to elicit model weights and alignment data. Anthropic claims the volume of anomalous API calls from Chinese IP addresses spiked by 1,800% over three months, correlating with the release of Alibaba’s Qwen3-110B model, which appears to share unusual feature representations with Claude 4.
For those who find the technical jargon dizzying, think of it this way. Imagine a master watchmaker guards a secret method for assembling a chronometer. A rival buys the watch, dismantles it with surgical precision, takes photographs of every gear and spring, and then builds a near-copy using slightly different materials. Anthropic alleges Alibaba did exactly that, but with software.
Alibaba has responded with characteristic defiance. In a statement from its Hangzhou headquarters, a spokesperson said: “These allegations are baseless and represent an attempt to stifle legitimate competition. Qwen3 is the product of years of independent research by thousands of our engineers.” The company has vowed to countersue, citing Anthropic’s reliance on open-source code for parts of its training infrastructure.
What makes this episode so consequential is not just the legal mudslinging, but the geopolitical context. Downing Street has seized the moment. In an unexpected development, the UK’s Department for Science, Innovation and Technology issued a joint statement with the National Cyber Security Centre warning of “state-aligned actors targeting UK-based AI labs”. The government is rushing through a new Digital Sovereignty Bill that would compel all foundational model providers operating in the UK to disclose foreign investment links and submit to “systematic red-teaming” of their supply chains.
The message from London is clear: the UK will not be a passive playground for tech giants battling over algorithms. As a senior Whitehall source put it, “We cannot have our national AI capabilities hollowed out by extraction operations dressed up as commercial research.” The bill, expected to pass before the summer recess, would create a statutory register of AI model training data provenance, a kind of nutritional label for AI systems.
Yet there is a deeper irony here that cannot be ignored. Anthropic and Alibaba both preach the gospel of responsible AI. Both have signed voluntary commitments on safety and transparency. But when the commercial stakes are this high, ethics become the first casualty. The same model alignment techniques designed to prevent AI from harming humans are now being weaponized in corporate espionage.
What happens next? This is a user experience question for society. The average person scrolling through a chatbot in Shanghai or London does not care whether the model’s weights were legally obtained. They care about the quality of the response. But this incident will force a reckoning. We will see regimes of digital sovereignty that fragment the global internet further. We will see AI models become black boxes not by design but by legal necessity. And we will see the cost of model training skyrocket as labs build impenetrable walls around their data.
Anthropic’s lawsuit is set for a preliminary hearing in July. Alibaba’s counterclaims will follow. But the real trial is unfolding in the court of public opinion and in the corridors of power. The UK’s move may inspire copycat legislation in Brussels, Canberra and Ottawa. The era of the AI commonwealth is giving way to fortified digital fiefdoms. And the user? They will be left with a more secure, but less interoperable, artificial intelligence.
