China escalates AI access dispute before Sept. 24 Xi-Trump summit

China has raised concerns that Anthropic's Claude Mythos could be weaponized, questioning why the US AI maker blocks Chinese access for normal purposes, according to people familiar with the matter.

"China still hopes to maintain a relatively friendly atmosphere before President Xi Jinping visits the US on Sept. 24, seeking a win-win outcome in AI talks expected before the leaders' meeting," Bloomberg reported, citing sources.

The concerns follow the US government's June decision to halt foreigners' access to Anthropic's Fable 5 and Mythos 5 models over national security risks. Anthropic disclosed in a June 10 letter to Congress that Chinese AI firm Alibaba executed the largest known distillation attack on Claude, using 25,000 fraudulent accounts and 28.8 million exchanges to extract roughly 57.6 billion tokens of digital knowledge — the equivalent of about 576,000 books.

The standoff places AI at the center of US-China tech decoupling. Anthropic's Mythos models demonstrated the ability to chain hundreds of complex software flaws and write functional exploit code on first attempt in more than 83 percent of cases, according to testing cited by cybersecurity analysts. That capability has driven Washington's push for export controls on model weights and high-end computing chips, while Beijing views access restrictions as a technology blockade.

The tension escalated after Anthropic's letter to Congress called for measures to "close loopholes allowing PRC AI labs to access advanced US chips" and "penalize PRC labs responsible for distillation attacks." The company said Alibaba's campaign targeted Claude's most valuable capabilities, including agentic reasoning, software engineering, and long-horizon tasks.

The distillation attack was notable for its scale and sophistication. Anthropic said the operation ran from April to June 2026, with each of the 25,000 fake accounts averaging roughly 13 exchanges per day — a volume low enough to evade detection. The extracted tokens represent less than 1 percent of what a major LLM typically trains on (10 to 15 trillion tokens), but the refined nature of the stolen knowledge makes it disproportionately valuable.

The Weaponization Question

China's concerns about weaponization reflect a broader anxiety about strategic asymmetry in AI capabilities. While US frontier models like Claude Mythos and OpenAI's GPT-5 have demonstrated advanced capabilities in software exploitation and autonomous reasoning, Chinese models have largely focused on commercial applications. The gap has narrowed in some areas — Alibaba's Qwen models and DeepSeek have shown competitive performance on certain benchmarks — but the gap in agentic and security-critical capabilities remains wide.

The US government's June action to restrict foreign access to Anthropic's most advanced models was unprecedented. Anthropic was forced to abruptly disable access globally for all customers because it could not immediately guarantee who was using the models. This created significant disruption for legitimate international users and highlighted the difficulty of enforcing access controls in a globally distributed AI market.

Sept. 24 Summit: What's at Stake

The AI talks expected before the Sept. 24 summit could determine the trajectory of AI governance between the world's two largest economies. Key issues include export control enforcement on advanced chips and model weights, data sharing and threat intelligence cooperation between AI labs, standards for AI safety, and access reciprocity for AI models and services.

For markets, the outcome carries direct implications. AI-sector stocks with China exposure face regulatory risk either way — tighter restrictions could hurt revenue growth for US AI companies serving Chinese customers, while looser controls could accelerate Chinese AI development and intensify competition. The last time the US imposed major AI-related export controls in October 2022, Nvidia's stock fell 8 percent in a single session before recovering as data center demand proved resilient. White House science adviser Michael Kratsios has separately outlined a policy framework for countering "adversarial distillation" of American AI models, treating the issue as a national security priority.

This article is for informational purposes only and does not constitute investment advice.