Anthropic just published a research paper that reads less like AI safety research and more like a classified intelligence briefing. The core message: America has a narrow window to maintain AI leadership over China, and current policy loopholes are closing it fast.
The paper outlines two dramatically different scenarios for 2028. In one, the US maintains a meaningful technological lead. In the other, China achieves near-parity and begins setting global norms for how AI is governed and deployed.
What makes this interesting is the mechanism. Anthropic argues that America's current advantage comes from compute - specifically, access to advanced chips from NVIDIA, TSMC, and ASML that Chinese companies can't replicate or legally acquire. Export controls have made that gap real.
But Chinese labs have stayed closer than they should through two workarounds that Anthropic characterizes as "industrial espionage."
First: chip smuggling and overseas data center access. Despite export controls, Chinese AI labs are apparently training models on restricted American hardware. A Supermicro co-founder was recently charged with diverting $2.5 billion worth of AI servers to China. That's not a rounding error - that's industrial-scale evasion.
Second: distillation attacks. Chinese researchers are creating thousands of fake accounts on US AI platforms, harvesting model outputs at massive scale, and using that data to train their own models. They're essentially free-riding on billions in American R&D, using the outputs to reverse-engineer competitive capabilities.
Anthropic's new model, Mythos Preview, apparently helped Firefox identify and fix more security vulnerabilities in one month than in all of 2025. That's the kind of capability jump the company warns shouldn't end up in Chinese government hands, particularly around autonomous vulnerability discovery that could be weaponized.
criminalize distillation attacks as industrial espionage, enforce export controls properly, close the overseas data center loophole. If the US doesn't act, Anthropic projects China reaches near-parity by 2028, floods global markets with cheaper models, and the CCP ends up shaping AI norms globally.
