How AI Labs Can Vet Foreign Talent Without Losing Them

The Trump administration's conflicting signals on foreign AI researchers have exposed a critical tension: how to safeguard national security without driving away the talent that built America's AI advantage.

In June 2026, the Commerce Department's Bureau of Industry and Security informed Anthropic that exporting its Claude Mythos 5 and Fable 5 models to foreign persons—including the company's own foreign employees in the United States—would require a license. Unable to screen users by nationality, Anthropic suspended global access for two weeks until Commerce lifted the controls. The episode revived industry fears of broader restrictions on foreign talent at U.S. labs, according to reporting by The Information.

Yet a week earlier, National Security Presidential Memorandum 11 had directed defense and intelligence agencies to build partnerships with AI companies, explicitly offering "assisting with personnel vetting" as one form of support.

Why it matters

Foreign-born or foreign-educated researchers comprise roughly 70 percent of leading U.S.-based AI talent, according to the Center for Security and Emerging Technology. Thirty-eight percent of top AI conference authors received undergraduate education in China. Policies that effectively bar foreign nationals from frontier work will push this talent to competitors—primarily China, which actively recruits through programs like Qiming and has set a national goal of achieving "competitive advantages in talent competition" by 2035.

The deemed-export challenge

Releasing controlled technology to a foreign person inside the United States counts as an export under U.S. regulations—a "deemed export" to that person's country of nationality. The risk intensifies when foreign employees access unreleased models, source code, or evaluation workflows that could generate export-controlled outputs. A pre-deployment chemical, biological, radiological, or nuclear evaluation, for instance, could release controlled information to whoever runs it.

Defense contractors, semiconductor manufacturers, and research universities have managed this problem for decades using technology control plans (TCPs)—documented procedures governing who may access controlled technology and under what conditions.

Risk-based vetting as the solution

A TCP for AI labs should include graduated personnel vetting that scales with access sensitivity. Most employees need only basic screening. Those querying unreleased models or running security evaluations warrant deeper review of "substantive contacts" with certain countries—regular travel, ongoing business ties, or other diversion risks. The few who could move raw model weights or disable safeguards require the most rigorous scrutiny, potentially adapting the government's own SEAD-4 adjudicative guidelines.

This approach mirrors how the government grades its own positions by potential damage and scales investigations accordingly. International Traffic in Arms Regulations already specify that nationality alone does not prohibit access to defense articles; instead, screening focuses on actual risk indicators.

Implementation obstacles

State employment laws, particularly California's Investigative Consumer Reporting Agencies Act, can complicate security vetting by requiring disclosure of investigation scope and providing employees the right to request copies of reports. If NSPM-11 results in federal vetting assistance, counterintelligence-sensitive information could remain with the government while labs receive access recommendations. Congress could also enact narrow preemption for defined sensitive AI roles.

Some cases will still require deemed-export licenses, particularly for Chinese nationals working with controlled material, as existing license exceptions cover only partner countries. The International Traffic in Arms Regulations present harder constraints, with a policy of denial for Chinese nationals that may require excluding certain employees from weapons-related evaluation work rather than model access generally.

The path forward

The Anthropic incident demonstrated that foreign-person model access can become an export control issue overnight. Rather than leaving labs to infer requirements from company-specific directives or effectively barring foreign talent, the government should publish a clear, risk-based vetting framework built on the TCP template agencies already use in other sectors.

Matching the vetting burden to actual risk—rather than treating every foreign national as a presumptive threat—would let U.S. labs compete for the world's best researchers while managing legitimate security concerns.

These details were first reported by Just Security.