A new front in the AI cold war has opened. Anthropic, the San Francisco-based AI safety company, has publicly accused Chinese e-commerce giant Alibaba of systematically extracting proprietary data from its Claude model. The allegation, which Alibaba has vehemently denied, has prompted British intelligence agencies to raise their threat level for technology theft.
According to sources within GCHQ, the UK’s signals intelligence agency, the extraction appears to have targeted the reinforcement learning from human feedback (RLHF) data that underpins Claude’s alignment. This is not a simple scraping of public outputs. This is a surgical operation, likely using a series of carefully crafted prompts to reverse-engineer the model’s safety guardrails and training methodology.
The implications are stark. If Alibaba has indeed succeeded, it could replicate Anthropic’s approach to constitutional AI, a framework designed to keep large language models honest and safe. For a company that has faced criticism over censorship and surveillance, the ability to embed its own ethical constraints into a powerful AI could be a tool for control as much as for innovation.
British intelligence is particularly concerned because Anthropic’s early safety research was partially funded by UK taxpayers through Innovate UK. The UK has positioned itself as a hub for AI safety, hosting the world’s first AI safety summit at Bletchley Park. A leak of this nature undermines that leadership and exposes vulnerabilities in the global AI supply chain.
Anthropic’s CEO, Dario Amodei, has called for an international investigation and tighter export controls on AI training data. In a statement, he said: “We have a moral imperative to ensure that the fruits of our labour are not weaponised by authoritarian regimes. This is not about competition. This is about existential risk.”
The accusation arrives at a delicate moment in US-China tech relations. The Biden administration has already imposed restrictions on the export of advanced AI chips to China. But data, unlike hardware, flows freely across borders. And as models become more powerful, the value of their hidden weights and training data rivals that of nuclear secrets.
Alibaba has responded with characteristic defiance. In a press release, the company stated: “These allegations are baseless and appear to be a cynical attempt to undermine a legitimate competitor. We operate in full compliance with international law and have no need to engage in industrial espionage.” No evidence has been publicly provided by either side.
The story is still developing. But what is clear is that the race to build safe AI is no longer just about who gets there first. It is about who gets to keep their secrets. For the British public, this means that the chatbot you use tomorrow might be the product of a stolen blueprint. And the intelligence community is watching.
My concern is that this is a glimpse of a darker future. As AI models become more integrated into critical infrastructure, from healthcare to defence, the extraction of their inner workings becomes a matter of national security. We are moving from a world of physical espionage to one of algorithmic heists.
Julian Vane, Technology & Innovation Lead, signing off.
