Why the White House is testing top-tier AI models

The White House has finalized a new, voluntary framework to test advanced AI before it is released to the public. Essentially, top AI companies can send their newest systems to the federal government for a 30-day security review. During this time, officials will use a private, classified system to check if the AI has dangerous capabilities, such as the ability to hack into other software or critical infrastructure. This plan currently only applies to the most powerful models, known as frontier models, and specifically excludes open-source AI—systems where the core building blocks are made available for anyone to download and modify. The government is keeping the specific details of its testing criteria a secret, a decision that has drawn criticism from researchers who believe the process should be transparent.

Think of an AI model like a very complex, highly trained apprentice. To build one, companies feed a massive dataset—a huge collection of text, code, or images—into a digital structure called a neural network. This network adjusts billions of small settings, known as parameters, until it learns how to predict the next piece of information in a sequence. Once this training is done, you have a finished model that can perform tasks on demand. Testing these models is difficult because they are essentially black boxes; they contain so many interconnected layers that even their creators do not always know exactly how they will behave in a novel situation. The government’s new review process is meant to act like an extra round of quality assurance. By forcing companies to hand over their models before they are released, officials hope to find hidden vulnerabilities that could be exploited by malicious actors, rather than discovering them only after the tool is already live and available to millions of users.

The secrecy surrounding this framework is the core of the debate. Because the government is not disclosing how it defines risks or which specific tests it uses, critics argue that the system is essentially a "trust us" agreement with big tech companies. If only the largest firms know the rules of the road, it may become harder for smaller startups to compete, effectively cementing the current industry leaders. As AI agents gain the ability to take actions independently—such as navigating websites or controlling software—the stakes for these security reviews grow. The government is trying to find a balance between encouraging innovation and preventing harm, but for now, they are doing so behind a closed door. The big question remains: can we rely on a private process to keep us safe, or is transparency necessary for the public to truly trust these systems?