A recent development in artificial intelligence has raised eyebrows in cybersecurity circles. Mistral AI, a firm based in Paris, has released an impressive trillion-parameter AI model titled Mistral Large 4, affectionately referred to as “Le Chonk.” The model is now freely available for download, modification, and execution, which means it can be utilized outside the developer’s direct control. However, concerns have emerged regarding its potential misuse by malicious actors looking to conduct real-world cyberattacks.
The release of this model follows weeks of public preview during which Mistral monitored its use, especially in cybersecurity contexts. According to the company, while the model effectively resisted black-box attacks against live targets during its testing phase, they acknowledged the potential for it to be exploited in malicious ways as capabilities continue to advance. In a blog post outlining the release, Mistral stated, “Though we observed the model refusing to comply with black-box attacks against live targets, we see a potential for uplift to conduct real attacks.” This alarming caveat suggests that the very strength of the model could also be its Achilles’ heel.
Mistral emphasizes its commitment to enhancing cybersecurity by modifying the model’s built-in guardrails and plans to continue updates in the future. However, these updates will only apply to the managed API and future versions, leaving those who have already downloaded the model without the benefit of ongoing improvements. Pierre Stock, the Vice President of Science at Mistral, explained that existing copies of the model would not receive automatic updates, underscoring the risks of deploying outdated versions.
In their ongoing research and monitoring, Mistral has noted that while the model had some success in aiding defensive efforts, it did not specify the attack stages where it could be helpful. Stock mentioned that Mistral Large 4 had exhibited the ability to assist in identifying and addressing software vulnerabilities, which is a crucial aspect of cybersecurity. However, he also emphasized their awareness of the potential for misuse as AI capabilities evolve, indicating a continual commitment to reinforcing safeguards.
Hugging Face, the platform hosting the Mistral weights, pointed out that the release of open weights significantly shifts both usage policies and deployment responsibilities to the end-users. This raises essential questions around security, especially considering a report from the UK’s AI Security Institute. Their findings indicate that prominent open-weight models typically lag behind closed systems in terms of their effectiveness on cyber tasks by several months. As they stated, the open release of highly capable models carries a “persistent and irreversible risk of misuse.”
The risks associated with the Mistral release are not hypothetical. Recently, Microsoft revealed that a group aligned with China utilized the agentic AI framework, ARTEX, in a sophisticated attack targeting more than 2,250 Microsoft 365 tenants. Such incidents highlight the vulnerabilities that can arise from the open availability of advanced AI tools.
Mistral maintains that their model can differentiate between legitimate defensive maneuvers and malicious requests. Interestingly, they claim that in the cyber-related assessments conducted, Mistral Large 4 rejected 96.6% of malicious prompts, a performance metric considered superior among open models. However, analysts caution against viewing this figure as a universal indicator of security. Niyati Daftary, a principal analyst at Gartner, cautioned that the 96.6% success rate applies only to Mistral’s test scenarios, which may not translate to real-world applications with varying conditions and adversarial use cases.
To address specific cybersecurity needs, Mistral is working with designated partners like Orange Cyberdefense to provide access to a specialized version of the model that is optimized for penetration testing and vulnerability assessments. These tailored versions are subject to strict vetting processes and contractual safeguards, ensuring they are used responsibly and only on authorized systems.
The company has also shared some of its benchmark scores. Mistral Large 4 achieved a 51% rating on the Artificial Analysis Cyber Index, a measurement tool for assessing how well an AI model identifies, reproduces, and patches vulnerabilities. Despite this achievement, some observers noted that two other open-weight models scored higher, calling into question Mistral’s claim of superiority.
In discussions about compliance and safety, Mistral’s Stock affirmed that the company adheres to the EU’s General-Purpose AI Code of Practice. However, specific details about whether Mistral Large 4 qualifies as a systemic-risk model under EU regulation have not been disclosed, leading to questions about the risk assessments undertaken prior to the model’s public release.
The implications of Mistral’s open-weight AI model for enterprises are significant. The model is released under a modified MIT license, allowing for various uses, including modification and commercialization. However, companies that intend to use it as part of a model-as-a-service offering will need a separate licensing agreement. For those opting not to manage the model themselves, Mistral offers an API priced at $0.70 per million input tokens and $2.10 per million output tokens.
Experts suggest that although sovereignty—keeping operations within local jurisdictions—can be a compelling factor for European buyers, it should not overshadow security concerns. As Charlie Dai from Forrester pointed out, open weights and self-hosting place the burden of responsibility for monitoring, access control, and misuse prevention on the organizations utilizing them.
In conclusion, while Mistral Large 4 presents a powerful tool for security work, it also opens the door to potential exploitation if not properly managed. As enterprises consider self-hosting this model for tasks like code audits or vulnerability triage, it’s crucial that they approach its deployment with stringent precautions, ensuring that it operates within controlled environments and with adequate human oversight.