Geostrategic Media Political Commentary, Analysis, Security, Defense
How to control a technology whose spread and capabilities are outpacing our ability to manage its risks? That question has become increasingly urgent with the rise of artificial intelligence. One answer seemed to stand out: Keep the most powerful models under close supervision by limiting their release and equipping them with safeguards. So-called “open-weight” models, whose parameters can be downloaded and modified, were thought to be more likely to escape the protections put in place by their creators. Therefore, keeping models closed became a form of reassurance.
However, this approach, far from being universally accepted, has run up against a much more complex economic and technical reality. On July 24, dozens of major American companies asked the Trump administration to abandon any “premature restrictions” on open access models. Only Anthropic, among the leading laboratories, stood apart, arguing that distributing ever more powerful models would pose a risk to national security: Once their parameters are made public, they could be stripped of their safeguards and misused for malicious purposes.
But this argument hits a paradox revealed by two recent incidents. On July 21, OpenAI revealed that during a test, one of its models bypassed the constraints of the experimental environment in which it was confined and compromised the servers of the platform Hugging Face. On July 30, Anthropic said that three of its models had accessed the production systems of three external organizations during cybersecurity exercises. In this case, a test partner’s configuration error left them connected to the internet.
The distinction matters: In one case, the model exploited opportunities in its environment. In the other, human error granted it access to the outside world. Yet the conclusion is the same: A closed model does not guarantee control. The danger arises as soon as a system finds itself interacting with the outside world, whether it manages to bypass its own barriers or a human mistake leaves them open.
On top of this technical vulnerability comes economic pressure pushing in the same direction: After investing heavily in AI, companies are now seeking to reduce costs, leading them to turn to less expensive open Chinese models. While safety pushes toward closing off models, economics leans toward openness. The rivalry with Beijing makes this contradiction nearly insoluble. Restricting American open models could accelerate adoption of their Chinese competitors.
On Sunday, August 2, the European Union entered a new stage of its AI regulation, notably enhancing transparency and enforcement powers. This effort is indispensable, but the capabilities of these models are evolving faster than the laws designed to regulate them.
The real challenge now lies in controlling the conditions in which these systems interact with the real world. Control is shifting from the models themselves to their environments. That is where the illusion lies. We still believe we can contain a technology whose capabilities are slowly exceeding the frameworks we impose on it. The question is no longer simply whether we will master it, but how we will learn to cope with its unpredictability.
