The U.S. government's decision to restrict foreign access to Anthropic's advanced models has backfired spectacularly, accelerating the rise of competitors rather than protecting American dominance. Four days after the export controls took effect, rival AI labs are capturing the momentum Washington intended to preserve for domestic players.

Cohere reports a flood of government inquiries following Anthropic's restrictions. DeepSeek, China's leading AI lab, closed a record $7.4 billion funding round—its largest ever. Chinese AI companies are slashing token prices by up to 99 percent, making their models dramatically cheaper than American alternatives.

The export control strategy reveals a critical flaw. By cutting off Anthropic's international revenue streams, the policy doesn't eliminate competition. It redirects capital and users toward foreign labs with fewer restrictions. DeepSeek's funding windfall suggests investors now see Chinese AI as a viable long-term alternative rather than a secondary option. Lower pricing accelerates adoption globally, particularly in developing markets where cost matters most.

The timing compounds the problem. Washington implemented controls to maintain American technical leadership, but the immediate effect is to make non-American options more attractive. Users and companies now have stronger incentives to build on DeepSeek, Cohere, and other alternatives rather than Anthropic's restricted models.

Meanwhile, a separate vulnerability undermines the entire AI supply chain. Researchers discovered 144 poisoned npm packages that harvest credentials from AI development environments. The attack exploits the fact that developers often leave authentication tokens exposed in code repositories. Bad actors insert malicious code into popular packages, which then steal those credentials when developers install updates.

This supply chain attack demonstrates that export controls alone don't secure AI development. The infrastructure supporting AI labs remains porous. Attackers can inject themselves into the development pipeline without needing to build competing models. They simply compromise the tools developers already