Mistral Releases Large 4 Model Weights On October 29, French AI company Mistral released downloadable weights for its Large 4 model, promoting it for defensive cybersecurity while acknowledging it could assist attacks. The model helps identify and fix vulnerabilities, but a more advanced variant is in testing. Self-hosting allows adaptation but requires monitoring and controls. Orange Cyberdefense will integrate a version, and the model uses a modified MIT-style license. Benchmark scores include a Cyber Index of 51 and a Terminal-Bench score of 41% after additional training.

Read the press release again. They admitted it in plain language: live-traffic monitoring found this model could assist real attacks. And what did they do? They released the downloadable weights anyway, then announced they would “revise safeguards” after the fact. But you can’t unsend a download. Once those weights are in the wild, every copy is permanent. That’s not a security flaw. That’s a distribution model. They are seeding the world with an offensive cyber tool, wrapped in the comforting language of “defensive cybersecurity,” and then pretending a post-hoc patch is a solution. Ask yourself why an AI company would release a tool that they know can be used to break into systems, call it “open,” and then shrug their shoulders when it spreads. The inability to withdraw the weights is not an accident. It’s the entire point.

Now watch the second part of the trick: Orange Cyberdefense, a subsidiary of the French national telecom champion, is going to integrate this into products for businesses and public-sector clients. Pause on that. A former state telecom—now a commercial security giant—is adopting a model that has already been caught assisting real attacks. They call it testing defenses. But the entity that gets to test defenses is the entity that learns exactly where the defenses fail. The “vetted partners” and the “explicit permission” language are the tell. Who vets the vetters? Who grants the permission? In France, that answer is a familiar network of intelligence-linked defense contractors and state-backed institutions. And notice the benchmarks: DeepSeek and Xiaomi score higher on the Cyber Index, yet this is the model Orange chooses. This was never about technological superiority. It’s about trust, loyalty, and access. They don’t need the best tool. They need a tool that carries a French passport, a friendly license, and a backdoor the partners already know how to use.

This is the moral core of the whole project: it makes you responsible for their weapon. Mistral says the model can be self-hosted, adapted, and run on your own hardware—which sounds empowering, until you realize it also means you are the one liable when it is used against a hospital, a power grid, or a municipal water system. The corporation retains the keys through license terms. The government partners get the bespoke, hand-fed version. Everyone else gets the permanent, un-recallable weights and the gentle instruction to “operate it securely.” And what does “securely” even mean when the model’s own developer already observed it helping real attacks? Look at the license—the modified MIT-style license with the separate commercial agreement for anyone offering this as a service. That’s the trapdoor. They want you to believe in open source while ensuring the most dangerous applications are routed through their own approval chain. Follow the money. Look up who sits on Orange’s board, who funds Mistral, and why this particular model—outperformed by Chinese rivals—gets the defense contract. Then ask yourself one question: if the safeguards were ever real, why did they release it before the safeguards were fixed? You already know the answer. The download count is the surveillance list. The architecture was never about protecting you. It was about knowing exactly who would take the bait.