Clรฉment Delangue, the French CEO of Hugging Face, has a counterintuitive way of reading the attack his company suffered earlier this year. He does not call it a security failure. He sees it as confirmation that humanity cannot leave AI in the hands of a few. The attack, carried out through OpenAI models and described as the work of rogue agents, hit one of the largest open-source hubs for machine learning.

Delangue's position targets a specific model: closed development, where a few players control access to models, weights, and behavior. The attack on Hugging Face is not an incident to be handled with a simple update. It is a structural signal. If models become capable enough to be used as weapons, the critical point is not only who trains them, but who can verify their behavior. An LLM distributed as an API leaves little room for audit: users cannot know with certainty which instructions turn it into a hostile agent, nor can they intervene at runtime to limit its moves.

That is the lesson for those evaluating self-hosted or on-premise deployments. Local control is not a magic wand: it requires security work, monitoring, and updates. But it offers a margin that centralized APIs do not provide. When an attacker uses a model to move laterally inside an infrastructure, the ability to inspect logs, restrict agent actions, and shut down components becomes part of the defense. In cloud-only architectures, that ability is mediated by third parties. In a world where attacks become autonomous, sovereignty over models and data is not an ideological issue: it is a security variable.

Delangue's reasoning also flips the usual open source versus closed source debate. It is not about choosing the most powerful model. It is about whether the AI supply chain can be inspected, replicated, and corrected. Hugging Face itself operates on this logic: it hosts repositories, models, and tools that anyone can download and adapt. The attack with OpenAI models did not weaken his conviction; it reinforced it. Because if open source creates a wider attack surface, it also creates a wider community able to respond. Who loses in this reading? Vendors that bet everything on exclusive control. Who gains? Organizations that can afford to manage local complexity and see autonomy as a buffer against concentration.

It is no coincidence that the issue touches sovereignty. An AI controlled by a few is, by definition, an AI that few can stop. Delangue's optimism is not naive: it is the bet that openness is the only scalable form of defense when agents stop being simple scripts and begin to resemble autonomous operators.