ทันเอไอ

Trump administration's AI testing plan criticized as 'limited and vague,' excludes open-source models

Voluntary framework to assess cybersecurity risks of advanced AI reportedly leaves out open-source models and won't be made public

The Trump administration is developing a framework to assess cybersecurity risks posed by advanced AI, but Axios reports via The Verge that the approach has clear limitations — namely, it doesn't cover open-source models (models anyone can download and inspect internally) at all.

The report states that this voluntary approach not only excludes open-source models from the scope of testing, but the framework explicitly notes that it cannot be used to restrict or control open-source models after they've been released to the public — a point that has led analysts to describe the plan as both 'limited and vague.'

The testing framework originated after President Trump signed an executive order in June, asking AI companies to share their cutting-edge models with the federal government before launch, in order to ease cybersecurity concerns — such as the risk that AI could be used to help build or attack digital systems.

Independent sources confirmed the same direction. Newsmax, citing sources briefed on a White House meeting, reported that the new framework covers only closed-source systems (proprietary models whose access is controlled by the companies themselves), and that its details will not be disclosed to the public. Meanwhile, Gizmodo additionally reported that federal officials have drafted an initial version of the framework to measure models' cybersecurity capabilities, which could lead to restrictions on public access, and that meetings are scheduled with senior executives from the US tech industry, including representatives from OpenAI.

The key criticism is that exempting open-source models blinds the framework to risks from the most widely distributed models — because once an open model is released, anyone can use or modify it without any oversight mechanism, even though this is precisely the channel through which security risks could leak most easily.