Anthropic Says One Chinese AI Lab Ran 151 Million Exchanges Trying to Extract Claude's Training Data. Its Own Report Confirms What Washington Alleged Two Days Earlier.

This time, the accusation isn't a government advisory — it's the target company's own disclosure.

Anthropic published its most detailed threat intelligence report to date on September 10, documenting misuse of its Claude models between December 2025 and August 2026 across seven categories of harm and more than 40 tracked threat groups. The report identifies seven Chinese AI labs — including Alibaba, DeepSeek, and Moonshot — running what Anthropic calls illicit distillation campaigns, routing live conversations through Claude to extract training data without consent. One campaign alone exceeded 151 million exchanges.

The finding lands exactly two days after this newsletter covered a joint advisory from the NSA, CISA, and the FBI naming six Chinese firms, including Moonshot, for nearly identical conduct. The same distillation allegation is now confirmed by both a US intelligence assessment and the private company whose own models were the target, within the same week — and just as Moonshot pursues a $50 billion Hong Kong IPO built substantially on the model this newsletter has already connected to both disclosures.

The report's cyber-operations section documents a Russian state-linked espionage campaign, tracked as GTG-20006, with tradecraft Anthropic says is consistent with the group publicly known as Midnight Blizzard, targeting more than 20 Ukrainian and European government, military, and diplomatic organizations. What made the operation notable wasn't the targeting itself — Russian espionage against these institutions is a long-documented pattern — but the automation: the group built AI-driven workflows that monitored whether their malware had been flagged by security products, then automatically modified and redeployed it, collapsing work that used to require a team of skilled operators into a workflow a single person could run.

None of the affected Chinese labs has publicly disputed Anthropic's specific figures as of this writing. What makes this report different from the government advisory it echoes is the source: a private company documenting the theft of its own product, quantifying it in the hundreds of millions of exchanges, at the exact moment one of the named labs is preparing to go public on the strength of what it built with that data. The two disclosures don't just corroborate each other — they compound the cost of ignoring either one.