2028 AI Scenario: Anthropic Warns of China Chip Smuggling and Distillation Attacks

Anthropic, the AI safety company behind the Claude model family, has released a startling new research paper that reads less like a technical safety document and more like a geopolitical intelligence briefing. The paper, which outlines two competing futures for global AI leadership by 2028, centers on a single, decisive variable: the compute gap between the United States and China.

The Compute Gap: Key Variable in the US China AI Race

According to the report, the US currently holds a meaningful lead in frontier AI development, primarily due to its control over advanced semiconductor supply chains. American and allied companies—NVIDIA, TSMC, ASML, and others—have built chip fabrication and design capabilities that China cannot yet replicate. Export controls imposed by the Biden and subsequent administrations have made that gap real and measurable.

Anthropic 2028 AI Scenario: Two Possible Futures

Anthropic's 2028 AI scenario paper presents two starkly different outcomes. In the 'good' scenario, the US closes loopholes in export controls, widening the compute gap to 11x. In the 'bad' scenario, China reaches near-parity, reshaping global AI norms.

China's Chip Smuggling and Distillation Attacks

But the Anthropic paper warns that China's AI labs have stayed surprisingly close through two alarming workarounds. The first is large-scale chip smuggling. The report cites the recent criminal charges against a Supermicro co-founder, who was accused of diverting $2.5 billion worth of servers to China in violation of export controls. This suggests that Beijing is willing to break international law to acquire the hardware it needs.

Distillation Attacks: Industrial Espionage

The second workaround is even more insidious: distillation attacks. Chinese labs are reportedly creating thousands of fake accounts on US AI platforms, harvesting model outputs at scale, and using that data to train their own competing models. Anthropic explicitly calls this practice 'industrial espionage' and is pushing for legislation to criminalize it. The company argues that these attacks allow China to effectively free-ride on billions of dollars in US research and development.

AI Export Controls: The Stakes

TechCrunch reports that Anthropic's new model, Mythos Preview, which was released to select partners in April, demonstrated an extraordinary capability in cybersecurity. According to the article, Firefox used Mythos to fix more security bugs in one month than in all of 2025. This capability jump—specifically around autonomous vulnerability discovery—is precisely the kind of breakthrough that Anthropic warns China should not be the first to achieve.

Implications for Global AI Leadership

The Anthropic paper presents two starkly different scenarios for 2028. In the first, which the company describes as the 'good' outcome, the US closes the loopholes in its export control regime, enforces restrictions properly, and the compute gap widens to 11x. In this scenario, US models stay 12 to 24 months ahead of Chinese competitors, and democracies set the norms for how AI is governed globally.

Scenario 1: Democratic AI Governance

In the 'good' scenario, US models lead by 12-24 months, and democratic values shape global AI norms. This outcome requires immediate policy action.

Scenario 2: Authoritarian AI Influence

In the second, 'bad' scenario, the US fails to act. China reaches near-parity with US labs, floods global markets with cheaper AI models, and the Chinese Communist Party ends up shaping global AI norms. This could include exporting AI-enabled surveillance tools to other authoritarian governments, fundamentally altering the balance of power in international affairs.

TechRadar confirms the cybersecurity angle, noting that Mozilla announced that Anthropic's Mythos Preview and other AI models helped it identify and ship 423 Firefox security bug fixes in just one month. WIRED separately reported that Mozilla used Mythos to find and fix 271 bugs in Firefox, calling it a breakthrough in automated vulnerability discovery.

Geopolitical Implications of AI

The paper's framing has already sparked debate. Anthropic is positioning itself as a political actor, not just an AI research lab. By calling distillation attacks 'industrial espionage' and pushing for criminal legislation, the company is wading into policy territory that some argue is inappropriate for a private AI developer. Others counter that the stakes are too high to leave to government alone.

The core question remains: is the compute gap as decisive as Anthropic claims, or can algorithmic innovation close it? The answer will determine not just the future of AI, but the future of global power itself. As the 2028 deadline approaches, the choices made today by policymakers in Washington and Beijing will echo for decades.