Anthropic Warns US Must Close Loopholes to Prevent Authoritarian AI Dominance

AnthropicAnthropic

· Updated

Anthropic released a policy paper outlining two potential futures for global AI leadership in 2028 based on the competition between the US and China. The company argues that democracies must act now to close compute loopholes and deter model distillation attacks to maintain a 12-24 month lead. Securing this advantage is necessary to ensure democratic values shape the rules and norms of transformative AI.

Anthropic published a policy paper detailing the geopolitical stakes of the AI competition between the US and China. While the US leads in frontier models, Chinese labs are narrowing the gap by exploiting export control loopholes and conducting distillation attacks—using model outputs to replicate capabilities at a fraction of the original research cost.

The company warns that 2026 is a breakaway opportunity to lock in a 12-24 month lead. This urgency follows the Mythos Preview release, which demonstrated a 20x increase in security bug fixes. This acceleration mirrors the scale of the Anthropic Amazon compute expansion, which provides the physical infrastructure needed to sustain a commanding lead over authoritarian regimes.

Anthropic advocates for closing smuggling loopholes, deterring distillation attacks through legislation, and promoting the export of the American AI stack. This signals a shift toward more aggressive monitoring of model usage to prevent IP theft.

View the full update on anthropic.com
Anthropic
Anthropic
@AnthropicAI
X

We've published a paper that explains our views on AI competition between the US and China. The US and democratic allies hold the lead in frontier AI today. Read more on what it’ll take to keep that lead: https://t.co/TgJBeodWYK

1kretweets5.9klikes
View on X

Still wondering? A few quick answers below.

Anthropic outlines two futures. In the first, the US and its allies maintain a commanding 12-24 month lead by tightening export controls and stopping distillation attacks. In the second, the US fails to act, allowing the Chinese Communist Party to achieve parity and use frontier AI to automate repression and gain military advantages.

A distillation attack occurs when a lab uses thousands of fraudulent accounts to harvest outputs from a competitor's frontier AI model. By using these outputs to train their own systems, they can replicate advanced capabilities at a fraction of the original research and development cost, effectively free-riding on billions of dollars in American investment.

While the US currently leads in compute and model intelligence, Anthropic warns that China is narrowing the gap. Chinese labs are bypassing export controls by smuggling chips and accessing offshore data centers. Additionally, they use American models to accelerate their own research, threatening to erase the democratic lead before transformative AI arrives.

Claude Mythos Preview is a frontier model released to select partners in April 2026. It serves as a wake-up call for AI acceleration, demonstrating the ability to fix security bugs at 20 times the average monthly rate. Anthropic uses this model to illustrate how quickly AI capabilities are approaching the level of transformative intelligence.

Anthropic recommends three primary actions: closing loopholes for smuggled chips and foreign data center access, deterring distillation attacks through legislation, and promoting the global export of the American AI stack. These steps aim to lower the compute ceiling for authoritarian regimes while ensuring democratic nations set the global norms for AI use.

Every HeadsUpAI update is written based on its original source and reviewed before it's published. Read our editorial standards →

Anthropic

See all AI news & updates from Anthropic →

Share this update