The Trump administration has finalised a voluntary framework for testing advanced AI models for cybersecurity risks, but it explicitly excludes open-source models, which can be freely downloaded and inspected. Axios reports the guidelines even state they cannot be used to restrict open models once released. The framework follows a June executive order requiring AI firms to share frontier models with the federal government before release, and companies including Anthropic, OpenAI and Google were briefed on it at the White House, though the administration does not plan to publish the details.

The framework gives the government a 30-day window to review new models, but only applies to closed-source systems deemed "state-of-the-art" and posing a "national security risk" — terms the guidelines fail to define. Compliance is voluntary, yet major labs such as OpenAI and Anthropic have been seeking clarity on how to avoid triggering restrictions, and the lack of clear definitions leaves considerable ambiguity that could particularly disadvantage smaller AI providers trying to stay in the administration's good graces.