Mistral Releases Large 4 Open Weights as Cybersecurity Risks Remain
Mistral AI has put the weights for its trillion-parameter Large 4 model in public hands, giving developers the freedom to run and modify it while leaving its safeguards outside Mistral’s ongoing control. That trade-off is especially visible in cybersecurity: the company says the model refused attacks against live targets during preview, yet could still help users conduct real attacks.
The Paris-based company released the weights Thursday, several weeks after opening a public preview of the multimodal model, which it calls “Le Chonk.” Mistral said it monitored cybersecurity-related use during that preview and saw potential for the model to improve attack capability. It is adjusting the model’s safeguards and plans further updates in the coming months.
Downloaded copies won’t get new safeguards
Safeguards built into the model’s weights ship with the downloadable checkpoint. Later changes, however, will apply to Mistral’s managed API and future model releases. “Subsequent safeguard updates are applied to our managed API and future model releases, rather than automatically updating downloaded checkpoints,” Pierre Stock, Mistral’s vice president of science, told CSO.
That distinction puts the burden on organizations running their own copy to evaluate and maintain its protections. Hugging Face, which hosts the weights, made a similar point in its launch post: with open weights, deployment choices and adherence to the usage policy sit with the user.
Mistral said Large 4 refused 96.6% of malicious prompts in the cybersecurity portions of JailbreakBench, StrongREJECT and AgentHarm. The company described that as its highest refusal rate among open models. Niyati Daftary, principal analyst at Gartner, cautioned that the score reflects a specific test setup, not how the model will respond to an organization’s data, production environment or targeted attacks.
Stock didn’t identify which stages of an attack the preview model had helped with. He said its clearest gains were in defensive tasks: finding, reproducing and fixing software vulnerabilities. Mistral said the model refused requests to attack live targets in preview, while acknowledging that its capabilities could be misused.
The company is offering a separate cybersecurity variant through preview access for design partners, including Orange Cyberdefense. That version is optimized for penetration testing, vulnerability assessment and red-teaming. Mistral says access requires rigorous vetting and contractual safeguards, and testing must be limited to systems the partner owns or is authorized to assess.
Performance claims await more detail
Large 4 scored 51% on the Artificial Analysis Cyber Index, which evaluates how well models find, reproduce and patch vulnerabilities. Mistral’s published chart shows DeepSeek V4.1 Flash and MiMo-V2.6-Pro scoring higher.
Mistral also calls Large 4 the leading open-weight coding model, citing results on Cognition’s FrontierCode 2.0 and Terminal-Bench 4.0. The company hasn’t yet released the technical report behind those claims. Charlie Dai, Forrester’s vice president and principal analyst, said that makes the ranking difficult to assess across different workloads. With frontier models changing quickly, he added, leadership can shift within weeks or months.
Stock said a forthcoming report will detail Mistral’s cybersecurity evaluation methods, refusal tests and findings from preview monitoring. The company also plans to share relevant adversarial testing results while protecting sensitive security information, but hasn’t set a publication date.
External assessments provide useful context for the release. The UK’s AI Security Institute reported in July that leading open-weight models lagged closed frontier models by four to seven months on cyber tasks. It described open release of highly capable models as creating “a persistent and irreversible risk of misuse.” And earlier this month, Microsoft said a China-aligned group used the ARTEX agentic AI framework to target more than 2,250 Microsoft 365 tenants.
Enterprise control comes with operational work
Mistral released Large 4 under a modified MIT license that allows use, modification, fine-tuning and commercialization. A separate license is needed to offer the model as a service. Organizations that prefer Mistral’s API pay $0.70 per million input tokens and $2.10 per million output tokens.
For enterprises, self-hosting can provide more control over data and operations, but it also transfers responsibility for access controls, monitoring and misuse prevention. Daftary recommends treating the model as a controlled tool for code audits or vulnerability triage, rather than an autonomous agent. Her suggested measures include isolated environments, least-privilege access to code and credentials, sandboxing generated code, and requiring human approval before applying fixes.
European buyers may value the ability to host an open model themselves, Dai said, but sovereignty needs to be weighed alongside security. The EU’s cybersecurity agency, ENISA, said it is testing Large 4 to find vulnerabilities, audit code and analyze defensive logs.
The European Commission says models trained with more than 1025 floating-point operations are presumed to carry systemic risk under the EU AI Act. Providers in that category must assess and mitigate risks, report serious incidents and maintain cybersecurity protections. Mistral hasn’t said whether Large 4 meets that threshold or described a risk assessment conducted before release. Stock said the company has signed the EU’s General-Purpose AI Code of Practice, including its safety and security chapter, but did not clarify the model’s classification.
Large 4 gives organizations a capable model they can adapt to their own infrastructure, with a practical cost: downloaded copies won’t inherit future safeguard updates. For enterprise teams, the release makes deployment design, monitoring and human review part of the model decision from day one.
