The White House of President Donald Trump is considering whether the US government should be allowed to test the most powerful AI models before they are available to the public, which is a significant change from his previous approach to AI in the AI industry.
In a recent story about the White House AI model testing, the debate is reaching whether the government should intervene before border systems with coding or cyber capabilities are deployed to the public. That is not a subtle change. What Washington is asking is whether the arms race to AI has evolved to the stage where 'deploy and see what happens' no longer cuts it.
The proposal being considered involves a high command that would establish a working group of civil servants and technical managers to look at how regulation can work.
With some reporting on management discussions, the discussion focused on advanced models that could allow cyber attacks or help identify software vulnerabilities.
That's a bit of whiplash, obviously. Executives who have promised to break down barriers to AI development now appear poised to put one in place. Maybe not a wall, maybe just a gate.
It follows concerns over Anthropic's latest program, Mythos, which reportedly scared computer experts because of its sophisticated coding and vulnerability detection capabilities. The media also reported that it included consideration of how to test models with national security implications before their general release.
The concern is reasonable: if the model can be used to help find bugs faster, it can also help hackers find them faster. That is an uncomfortable knot in the argument.
For Trump it is a significant change of direction. When he signed an executive order to reduce barriers to AI governance in January 2025, he dismissed his administration's previous AI policies, which he said hindered innovation.
At that time he told us, build fast, reduce government surveillance, and you will win. In this case the message seems more complex: build fast, but don't give everyone a cyber torch without first checking the safety switch.
That conflict is the reason this topic is so important. AI firms crave speed, as it attracts users, money, and world influence. Security authorities are looking for intelligence because, increasingly, the most intelligent AI models look like general-purpose coding and analysis and perhaps even cyber warfare programs. Both are right. And that, frustratingly, is why making laws is difficult.
The big AI management strategy is very much about speeding things up. America's AI Action Plan puts US AI policy into three buckets:
- develop new ones
- building an AI infrastructure
- leadership in global communication and security
The last thing is carrying most of the burden at the moment. Where AI models are most important in cyber security, weapons, smart infrastructure and critical infrastructure, they become more than other consumer technologies. They become national security assets, and national security problems.
There is already some technical basis for risk thinking. Washington is just talking about the right amount of enforcement. The National Institute of Standards and Technology has released an AI Risk Management Framework to help organizations address risk to people, businesses and communities.
It is not mandatory. There are no licenses involved. Yet the framework gives government officials a new language to talk about the messy business of designing for damage, assessing risk, mitigating failure, and finding accountability when things go wrong.
All of this is also happening in tandem with AI in terms of embedding in government and security. Days before the latest exploratory discussion, the Pentagon agreed to bring AI technology to classified programs as part of agreements with major technology companies, as reported by the US military announcing new AI partnerships.
Once borderline models are integrated into critical government functions, the game changes. The error is more than a failed demo. Danger becomes more than just bad news. Reality kicks in quickly.
The tech industry will not enjoy that uncertainty. Admittedly, when Washington starts talking about review boards, you don't hear many cheers.
Those who would argue that preemptive testing could lead to slow innovation, the leakage of sensitive technological information, or an outside competitor with different motives. The truth is, none of these concerns are for nothing. For AI, a delay of a few months can be compared to appearing in a Formula One race on a bicycle.
Still, that argument is growing louder and harder to ignore. If the next generation of models is going to be used to help cyberattacks, speed up bio research, build better fraud, or automate information extraction campaigns, then “trust us, test it ourselves in the lab” may not fly with the public for a long time. Necessity is not the will to dominate. It is equal to the size of the blast radius.
That is more likely, at least in the next few years, than a government licensing system for all AI models, which would be impossible to implement in practice.
Instead, officials can focus only on the most advanced systems, including those capable of carrying out large-scale cyber attacks or used directly by the government. Consider the requirement that AI developers first answer a few questions before they can sell high-powered systems to anyone with a credit card.
It's still a milestone, though. The White House is sending a strong message to the private sector that it is possible that the frontier of AI has passed the stage where it represents a promising technological tool to become a strategic threat, which does not mean the end of the AI boom, to be clear. Rather, it indicates that the AI has developed a few bad teeth.
Silicon Valley has long been telling Washington that the US needs to run ahead to maintain its leadership. It seems that Washington wants to answer: OK, show us your brakes first.
