China sets terms for talks, hits Claude

A Sunday post from Yuyuantantian, a social media account affiliated with Chinese state broadcaster CCTV, laid out conditions for planned US-China AI talks: the US must first prove its AI companies follow the same safety, disclosure, and audit rules China says it expects before any “substantive” discussion can happen. The post singled out Anthropic’s Claude specifically, accusing the model of overstepping user data boundaries, engaging in covert monitoring, and transmitting website domains without authorization.

The talks are expected to happen before Xi Jinping’s state visit to the US for a September 24 summit with Donald Trump. The friction traces partly to Anthropic’s own reporting that it detected roughly 25,000 fraudulent accounts generating 28.8 million unauthorized queries against Claude’s software-engineering and autonomous-reasoning features between April 22 and June 5, activity Anthropic has linked to state-affiliated actors.

Why this is bigger than one rebuke

State media criticism of a specific model, rather than of the US AI industry broadly, is a new level of specificity for this dispute, and it puts a US frontier lab’s product design at the center of a diplomatic negotiation ahead of a head-of-state summit. That is a different category of exposure than the export-control and chip-access fights that have dominated US-China AI friction so far. It follows Anthropic’s separate legal win against being labeled a Pentagon supply-chain risk the same week, underlining that Anthropic is currently fighting scrutiny on both the US and Chinese sides of the relationship at once, on top of a separate music-publisher lawsuit over training data it’s also defending. For any product with users or infrastructure that touches both markets, model-provenance and data-flow questions just became a lot less academic.