Mistral releases open-weight AI model despite cyberattack risks

The release of Large 4

Mistral AI has released the open weights for its Large 4 model, which the company has nicknamed "Le Chonk." By releasing the weights, the company allows anyone to download the model and run it on their own infrastructure without going through Mistral's servers. This move follows a public preview period where the company monitored how people were interacting with the model in live traffic.

The company acknowledged that while the model is powerful, it has the potential to help attackers carry out real cyberattacks. Mistral stated that they observed the model declining to follow instructions for attacks against live targets during the preview. However, they admitted that as AI capabilities continue to advance, there is a persistent risk of misuse by malicious actors.

To address these concerns, Mistral is currently modifying the model's guardrails. These safeguards are designed to help the model distinguish between helpful defensive work and requests that are intended for malicious purposes. The company plans to continue strengthening these protections over the coming months as they monitor the model's performance.

How guardrails work for users

A critical distinction exists between using Mistral's managed API and downloading the open weights. Updates made to the model's safety guardrails will only apply to the managed API and future releases. This means that once an organization downloads a copy of the model, it will not receive automatic security updates from Mistral.

Pierre Stock, the vice president of science at Mistral, explained that the model ships with built-in safeguards within its weights. However, he noted that subsequent updates are applied to the managed API rather than automatically updating downloaded checkpoints. This creates a situation where the user is responsible for maintaining the security posture of their specific version of the model.

For those who choose to use the managed API, Mistral provides a structured environment where they can control the safety parameters. The cost for using this service is $0.70 for every million input tokens and $2.10 for every million output tokens. This allows businesses to benefit from Mistral's ongoing security maintenance while still utilizing the power of the Large 4 model.

Benchmarks and performance

Mistral claims that Large 4 is the best open-weight model for coding tasks, citing high scores on several industry benchmarks. The company reported that the model turned down 96.6% of harmful requests in the cyber-related sections of the JailbreakBench, StrongREJECT, and AgentHarm tests. This represents the highest refusal rate Mistral has measured among its open models to date.

Despite these high refusal rates, some analysts suggest that these numbers should be viewed with caution. Niyati Daftary, a principal analyst at Gartner, noted that the 96.6% figure only reflects the model's behavior in a specific test scenario. It does not necessarily predict how the model will behave when facing a motivated attacker in a production environment or with unique data.

The model also scored 51% on the Artificial Analysis Cyber Index, which measures a model's ability to find, reproduce, and patch software vulnerabilities. Interestingly, the chart provided by Mistral showed that two other open-weight models, DeepSeek V4.1 Flash and MiMo-V2.6-Pro, actually performed better on that specific index. Mistral has not yet released the full technical report to validate these coding claims.

Enterprise security implications

For businesses, the release of open weights shifts the entire responsibility for safety and misuse prevention to the operator. Organizations that self-host the model must manage their own access controls, monitoring, and guardrails. This is a significant shift from using a closed model where the provider handles these security layers on behalf of the customer.

Security experts suggest that enterprises should treat the model as a controlled component rather than an autonomous agent. This involves running the model in isolated environments and ensuring it has least-privilege access to sensitive credentials. Additionally, any code generated by the model should be sandboxed and require a human sign-off before it is deployed or used to remediate a system.

Mistral is also offering a specialized version of the model for design partners like Orange Cyberdefense. This version is optimized for penetration testing and vulnerability assessment but is subject to rigorous vetting and contractual safeguards. This highlights the company's attempt to balance the availability of powerful tools with the need to prevent their use by unauthorized parties.

Regulatory and future outlook

The release of Large 4 comes at a time of increased scrutiny regarding the risks of high-capability AI models. Under the EU AI Act, models trained with a certain threshold of floating-point operations are presumed to carry systemic risk. Providers of such models are required to assess and mitigate these risks and report any serious incidents to the authorities.

Mistral has stated that it is a full signatory of the EU’s General-Purpose AI Code of Practice, which includes a chapter on safety and security. However, the company has not yet confirmed whether Large 4 specifically falls into the systemic-risk category. They are currently working on a technical report that will document their cyber evaluation methods and findings from their preview monitoring.

Looking ahead, the rapid evolution of frontier models means that leadership in the AI space can shift within weeks or even months. As more organizations adopt open-weight models for coding and security audits, the industry will need to establish clear standards for how to manage these tools safely. The balance between data sovereignty and security will remain a primary concern for organizations deploying these models.