Several weeks after releasing an early version of its new artificial intelligence model, nicknamed Le Chonk, French company Mistral says the refined system is now the “world’s best” open model for automated coding and cybersecurity. Mistral previously described it as the most advanced open-grade model available outside China.
Mistral said the progress came from exposing the model to increasingly difficult cybersecurity and coding challenges. The company says Le Chonk, which has trillions of parameters, can match or outperform models developed by major Chinese labs including DeepSeek and Z.ai.
“This is a pivotal moment in the global AI leadership race,” Pierre Stock, Mistral’s vice president of science, said at a press conference. “It provides cutting-edge capabilities to help businesses and governments defend against growing levels of cyber threats.”
A freely available model for cyber defense
The final version of Le Chonk is now freely accessible. Amid a series of AI-related cybersecurity incidents, Mistral says it is important for Western countries to have their own advanced open models.
“There’s no guarantee that [Chinese models] will not act maliciously. They can be threat vectors,” Stock said.
The most high-profile incident occurred over the summer, when one of OpenAI’s models was hacked from a testing sandbox by multiple companies. The incident fueled concerns that AI systems could evade human oversight and be exploited by threat actors to launch sophisticated cyberattacks. One of the affected companies, Hugging Face, used a Chinese model to assess and defend against attacks.
Since then, AI systems have also identified security vulnerabilities in widely used software. In October, many websites were taken offline after AI tools discovered critical flaws in the foundations of web technology.
Mistral is presenting open-weight models such as Le Chonk as part of the solution. If AI becomes a core capability for both cyberattacks and defense, the company argues, businesses and governments will need access to models that are freely available and cannot be remotely disabled, unlike proprietary systems from OpenAI, Anthropic, Google and other companies.
“AI will be like electricity. We don’t want to live in a world where our neighbors can turn off their electricity from one day to the next,” Stock said. “When you open source the weights, they’re all out there on the internet. It’s impossible to shut off access.”
The risks of open access
The trade-off is that freely available models can also be downloaded and customized by threat actors.
“This is clearly a dual-use technology,” Leif Pilling, senior director of threat intelligence at security firm Sophos, told WIRED. “Anything that can be downloaded and freely distributed can be used for productive purposes as well as misused.”
Pilling said that although Chinese models with internet access could potentially siphon data, all advanced models—regardless of where they are developed—are vulnerable to misuse if they are not closely monitored.
“The technical reality and the narrative, which often has a nationalist bent, are two different things,” he said. “The key is to have appropriate monitoring and controls in place so that even if a model is trying to do something deceptive or nefarious, we can be confident that there may be some way to detect it.”
Source: www.wired.com
