Mistral Releases Large 4 Weights Amid Cybersecurity Concerns

Mistral released the weights for its 1-trillion-parameter multimodal model, Mistral Large 4, on Thursday, weeks after beginning a public preview. The company said anyone can download the model for free, modify it and run it in their own environment. It also gave the model the nickname “Le Chonk.” Mistral said it observed cybersecurity-related use during real traffic in the preview. The model refused requests for black-box attacks targeting real-world entities, but the company also said it could be misused to improve the ability to carry out real attacks.

Mistral plans to revise and update its guardrails over the coming months, but those updates will not apply to models already downloaded. Pierre Stock, Mistral’s vice president of science, said safeguards are built into the weights, while later changes will be delivered through its managed API and future model releases rather than automatically updating downloaded checkpoints. Asked which stages of attacks the model supported in preview traffic, Mistral gave no specific examples and said it had been effective at defensive work such as identifying, reproducing and fixing software vulnerabilities.

Hugging Face, which hosts the weights, said users will control usage policies and deployment after release. Mistral said Large 4 rejected 96.6% of malicious prompts across the cybersecurity sections of JailbreakBench, StrongREJECT and AgentHarm, the highest rate it had measured among open models. Gartner senior analyst Niyati Daffara said the figure reflects Mistral’s own tests and does not show how the model behaves with enterprise data or in real operating environments.

Mistral is giving partners, including Orange Cyberdefense, a preview of a cybersecurity-specific version tailored to penetration testing, vulnerability assessments and red-team work. Access is subject to strict screening and contractual safeguards, and the company said partners may test only systems they own or are authorized to assess. The UK AI Security Institute reported in July that leading open-weight models lagged state-of-the-art closed models by 4–7 months on cybersecurity tasks, and warned that releasing highly capable models poses persistent, irreversible misuse risks. Earlier this month, Microsoft said a China-linked hacking group targeted more than 2,250 Microsoft 365 tenants using the agentic AI framework ARTEX.

Large 4 scored 51% on the Artificial Analysis Cyber Index, which assesses vulnerability detection, reproduction and patching. A chart in Mistral’s post showed the open-weight models DeepSeek V4.1 Flash and MiMo-V2.6-Pro scoring higher. Mistral also called Large 4 the best open-weight model for coding, citing leading results on Cognition’s FrontierCode 2.0 and TerminalBench 4.0, but it has not published a technical report to support that claim. Charlie Dai, a vice president and principal analyst at Forrester, said the lack of a detailed report makes the claim difficult to verify across different work environments, and that leadership can change within months or weeks. Stock said a report would cover cybersecurity evaluation methods and results, tests of refusals for harmful requests, and preview monitoring; Mistral would also share adversarial testing results while protecting sensitive information. He gave no publication date.

The European Union Agency for Cybersecurity said it is testing Large 4 for vulnerability identification, code audits and defensive log analysis. The EU AI Act presumes a model poses systemic risk if the computing used to train it exceeds 10^25 FLOPs, and requires providers of such models to assess and mitigate risks, report serious incidents and implement cybersecurity protections. Mistral’s post did not say whether Large 4 falls into this category or what risk assessment it conducted before release. Stock said Mistral signed the full EU General-Purpose AI Code of Practice, including its safety and security chapter. The European Commission recognizes compliance with the code as one way to demonstrate fulfillment of legal obligations, but Large 4’s systemic-risk classification remains unconfirmed.

Mistral released the weights under a modified MIT license that permits use, modification, fine-tuning and commercialization. Businesses offering the model as a service must obtain a separate license. Companies that do not want to host it themselves can use Mistral’s API, priced at $0.70 per 1 million input tokens and $2.10 per 1 million output tokens. Dai said sovereignty is a meaningful differentiator for European companies but should not take precedence over security concerns; releasing weights and enabling self-hosting shift responsibility for guardrails, access controls, monitoring and misuse prevention to operators. Daffara advised companies hosting the model to run it as a controllable component rather than an autonomous agent, with an isolated environment, least-privilege access, sandboxed code execution and human approval before changes. Self-hosting

Mistral released the weights for its 1-trillion-parameter multimodal model, Mistral Large 4, on Thursday, weeks after beginning a public preview. The company said anyone can download the model for free, modify it and run it in their own environment. It also gave the model the nickname “Le Chonk.” Mistral said it observed cybersecurity-related use during real traffic in the preview. The model refused requests for black-box attacks targeting real-world entities, but the company also said it could be misused to improve the ability to carry out real attacks.

Mistral plans to revise and update its guardrails over the coming months, but those updates will not apply to models already downloaded. Pierre Stock, Mistral’s vice president of science, said safeguards are built into the weights, while later changes will be delivered through its managed API and future model releases rather than automatically updating downloaded checkpoints. Asked which stages of attacks the model supported in preview traffic, Mistral gave no specific examples and said it had been effective at defensive work such as identifying, reproducing and fixing software vulnerabilities.

Hugging Face, which hosts the weights, said users will control usage policies and deployment after release. Mistral said Large 4 rejected 96.6% of malicious prompts across the cybersecurity sections of JailbreakBench, StrongREJECT and AgentHarm, the highest rate it had measured among open models. Gartner senior analyst Niyati Daffara said the figure reflects Mistral’s own tests and does not show how the model behaves with enterprise data or in real operating environments.

Mistral is giving partners, including Orange Cyberdefense, a preview of a cybersecurity-specific version tailored to penetration testing, vulnerability assessments and red-team work. Access is subject to strict screening and contractual safeguards, and the company said partners may test only systems they own or are authorized to assess. The UK AI Security Institute reported in July that leading open-weight models lagged state-of-the-art closed models by 4–7 months on cybersecurity tasks, and warned that releasing highly capable models poses persistent, irreversible misuse risks. Earlier this month, Microsoft said a China-linked hacking group targeted more than 2,250 Microsoft 365 tenants using the agentic AI framework ARTEX.

Large 4 scored 51% on the Artificial Analysis Cyber Index, which assesses vulnerability detection, reproduction and patching. A chart in Mistral’s post showed the open-weight models DeepSeek V4.1 Flash and MiMo-V2.6-Pro scoring higher. Mistral also called Large 4 the best open-weight model for coding, citing leading results on Cognition’s FrontierCode 2.0 and TerminalBench 4.0, but it has not published a technical report to support that claim. Charlie Dai, a vice president and principal analyst at Forrester, said the lack of a detailed report makes the claim difficult to verify across different work environments, and that leadership can change within months or weeks. Stock said a report would cover cybersecurity evaluation methods and results, tests of refusals for harmful requests, and preview monitoring; Mistral would also share adversarial testing results while protecting sensitive information. He gave no publication date.

The European Union Agency for Cybersecurity said it is testing Large 4 for vulnerability identification, code audits and defensive log analysis. The EU AI Act presumes a model poses systemic risk if the computing used to train it exceeds 10^25 FLOPs, and requires providers of such models to assess and mitigate risks, report serious incidents and implement cybersecurity protections. Mistral’s post did not say whether Large 4 falls into this category or what risk assessment it conducted before release. Stock said Mistral signed the full EU General-Purpose AI Code of Practice, including its safety and security chapter. The European Commission recognizes compliance with the code as one way to demonstrate fulfillment of legal obligations, but Large 4’s systemic-risk classification remains unconfirmed.

Mistral released the weights under a modified MIT license that permits use, modification, fine-tuning and commercialization. Businesses offering the model as a service must obtain a separate license. Companies that do not want to host it themselves can use Mistral’s API, priced at $0.70 per 1 million input tokens and $2.10 per 1 million output tokens. Dai said sovereignty is a meaningful differentiator for European companies but should not take precedence over security concerns; releasing weights and enabling self-hosting shift responsibility for guardrails, access controls, monitoring and misuse prevention to operators. Daffara advised companies hosting the model to run it as a controllable component rather than an autonomous agent, with an isolated environment, least-privilege access, sandboxed code execution and human approval before changes. Self-hosting