Trump administration's AI safety framework sidesteps open models, lacks clear definitions

The Trump administration has introduced a framework for evaluating potential cybersecurity threats from cutting-edge AI systems, but the approach has significant blind spots. The voluntary guidelines deliberately exclude open-source models—software that anyone can freely download and examine—and explicitly bar the government from controlling such models once they're publicly available.

What the framework covers#

The testing plan applies only to closed-source AI systems that represent state-of-the-art capabilities and pose what officials deem national security risks. Companies participating in the voluntary program would have a 30-day window for the federal government to review their models before release. Major AI developers including OpenAI, Anthropic, and Google attended a White House briefing on the framework, though the administration has not committed to making the details public.

The problem: vague terms and unclear boundaries#

The framework's most glaring weakness is its failure to define critical terms. What qualifies as "state-of-the-art"? What exactly constitutes a "national security risk"? These absences undermine the initiative's stated purpose of assessing frontier AI safety. The ambiguity creates a particularly thorny situation for smaller AI companies seeking to understand—and comply with—government expectations without inadvertently crossing lines that remain undefined.

Why voluntary compliance matters#

Frontier AI labs like OpenAI and Anthropic have sought government guidance precisely because they want to avoid triggering formal restrictions. The vagueness of the framework means this mutual uncertainty will likely persist. Without clearer definitions and criteria, companies face genuine difficulty anticipating which systems might trigger regulatory scrutiny, potentially creating a chilling effect on development or conversely having little practical impact.