Mistral Gives Away a Trillion-Parameter Cyber Model and Admits Attackers Could Use It Too
Mistral released the open weights of its trillion-parameter Large 4, pitching it at cyber defenders while conceding in its own words that it could give attackers uplift. The company's benchmarks look strong, but a downloadable model cannot be recalled.
Mistral just put a trillion-parameter model on Hugging Face and told the world it is built for cyber defenders. In the same breath, it admitted the thing could help attackers too.
The French lab released the weights of Mistral Large 4, nicknamed Le Chonk, on Thursday. It is a multimodal model that Mistral first previewed earlier this month, and anyone can now download it. Closed labs are locking their strongest security models behind vetting. Mistral is betting the opposite way, and says so openly in its launch post.
Pierre Stock, Mistral's vice president of science, told TechCrunch that open weights give companies and governments frontier-grade defense they can run on their own hardware.
"Access that can't be cut off by definition because the model is open-weighted." Pierre Stock, Mistral vice president of science
That argument has teeth. A security team in the middle of an incident does not want an API key revoked or a filter that mistakes malware analysis for malware writing. Hugging Face's own write-up of the release makes the same case, noting that vulnerability research and incident response often look like attacks to a model's safety filters.
The numbers are Mistral's, so read them that way
Mistral says Large 4 scores 51 percent on the Artificial Analysis Cyber Index, which tests whether a model can find vulnerabilities in a codebase, reproduce them and patch them without breaking anything. On the reproduce-and-patch piece alone, Hugging Face's blog cites 82 percent, and says Mistral reports that several closed frontier models score near zero there because they refuse the task. The same post lists 93 percent on a set of 40 Cybench challenges. These are company-reported figures, and the 82 percent is a sub-test result, not the headline index score.
On coding, Mistral claims a 48 percent score on the Artificial Analysis Coding Agent Index, effectively tied with China's open-weight GLM-5.3 while using 47 percent fewer tokens per task. Hugging Face adds that, as of Thursday, Large 4 ranks fifth among open-weight models on Artificial Analysis' overall Intelligence Index.
Booz Allen, quoted by Mistral, says the model matched other frontier cyber models on vulnerability-detection recall at a better cost. Cisco's Amy Chang, also quoted there, notes that open weights let researchers study how models behave, including where they could be misused.
Now the awkward part. Mistral says it watched cyber-related use in live traffic after the preview and saw the model refusing black-box attacks on live targets. It still wrote this:
"We see a potential for uplift to conduct real attacks." Mistral, launch post
The company says it is modifying its guardrails and will keep updating them. It also claims an average refusal rate of 96.6 percent on malicious cyber requests, the highest it measures among open models, drawn from the cyber prompts in JailbreakBench, StrongREJECT and AgentHarm. Flip that around and about 3 in 100 such prompts do not get refused, on Mistral's own average.
The bigger issue is what a guardrail is on a model you can download. Hugging Face says open weights put the usage policy under the user's control. That cuts both ways. A defender can tune the model to stop refusing. So can anyone else, and no update from Mistral reaches a copy already sitting on someone's server.
Mistral seems to know it. Its more aggressive variant, built for penetration testing and red-teaming, goes only to vetted enterprise design partners such as Orange Cyberdefense, under contractual safeguards. When a company gates its sharpest version while releasing the general one, it is telling you where it draws the line.
Commercially, the license is the Mistral Ecosystem Open-Weight License, a modified MIT. Use, modify and commercialize freely, except that businesses selling the model as a service need a separate deal with Mistral.
This is a big swing for the Paris lab, whose funding round we covered earlier. Whether open cyber capability tips toward defenders or attackers is an open empirical question. Mistral's technical report is still coming, and independent testers now have the weights to check its claims.
