In a dramatic escalation of Western-Chinese tech tensions, Anthropic, the San Francisco-based AI safety company founded by former OpenAI researchers, has formally accused Alibaba of orchestrating a sophisticated, large-scale extraction of its proprietary machine learning models. The allegations, detailed in a confidential submission to the UK’s National Cyber Security Centre (NCSC), claim that operatives linked to the Chinese e-commerce giant systematically scraped Anthropic’s cloud APIs, reverse-engineered its reward models, and deployed adversarial queries to reconstruct the architecture of its flagship Claude model. British cybersecurity chiefs have now launched an urgent investigation, warning that downstream exposure could compromise UK firms relying on AI-driven decision-making in critical sectors, from financial compliance to medical diagnostics.
The accusation marks the first time a major Western AI lab has publicly accused a Chinese competitor of industrial espionage via algorithmic extraction. Anthropic’s evidence reportedly includes anomalous traffic patterns, timestamps synchronised with Beijing’s business hours, and IP addresses that resolve to Alibaba’s cloud subsidiary. In a statement to the press, Anthropic’s head of security, Dr Elena Voss, said: "We detected a sustained, multipronged attack over three months. The perpetrators didn’t just steal weights; they lifted entire decision trees and safety classifiers. This isn’t simple theft. It’s replication of the safety infrastructure we spent years building." The NCSC, which has been quietly probing the matter since May, expressed "deep concern" over the potential for UK universities and start-ups to become unwitting conduits for AI infiltration. "We are seeing a growing pattern where adversarial nation-states use commercial entities as proxies to strip Western AI of its protective layers," a senior official told The Guardian, speaking on condition of anonymity.
Alibaba has vehemently denied the allegations, issuing a statement calling them "baseless and defamatory" and accusing Anthropic of "resorting to protectionist smear tactics as their market lead erodes." The company’s AI division, Damo Academy, emphasised that all its models are developed through "open research and lawful reverse engineering of publicly available information." Yet the timing is awkward for Alibaba, which recently unveiled a suite of open-source models that suspiciously mirror certain Claude capabilities, including a preference-tuning mechanism that Anthropic claims is uniquely signatured. Industry analysts are divided: some view the accusation as a desperate move by a company losing ground to open-source alternatives, while others see it as a wake-up call about the fragility of frontier AI insights.
The British government’s interest reflects a broader shift in digital sovereignty concerns. Whitehall sources indicate that existing cybersecurity frameworks, designed to protect static data, are ill-equipped for AI value chains where knowledge can be exfiltrated through queries alone. The NCSC investigation will focus on whether any UK-based clients of Anthropic or Alibaba were compromised, and whether the extracted models could be weaponised to target British businesses. "This is Black Mirror brought to life," said Julian Vane, Technology & Innovation Lead at a London-based think tank. "We’re seeing the ultimate user experience nightmare: your algorithmic thoughts get sucked out and turned into a competitor’s product. But the real worry is when those cloned models start making decisions that affect people’s livelihoods without the safety guardrails." The probe is expected to conclude by autumn, but its findings could reshape export controls on AI model weights and ignite a global debate about the ethics of intelligence extraction in the machine learning age.
