Through a new partnership, Orange Cyberdefense is gaining access to a new variant of Mistral Large 4 designed for advanced security tasks. The initial initiatives will focus on penetration testing of IT systems and post-incident analysis and investigation.
Mistral AI today published Mistral Large 4 on Hugging Face, which the French AI company describes as the most powerful open model for agentic coding and cyber defense. Orange Cyberdefense also announced that it is the first European managed security service provider (MSSP) to serve as a design partner for a new variant of the model.
According to Mistral, Large 4 achieves a score of 93 percent on CyberBench, and the model is able to reproduce and patch 82 percent of real-world vulnerabilities. The training took place on 3,800 Nvidia Grace Blackwell GPUs in European data centers. While it may not be on par with U.S. AI labs, Mistral is once again somewhat close to the performance of competing open-weight model developers from China. More importantly, it is by far Europe’s best LLM. For a European security company like Orange Cyberdefense, Mistral is a top-tier option that has regained its competitive edge since the launch of Large 4.
Penetration testing and forensic investigation
Orange Cyberdefense plans to gradually integrate AI systems based on Large 4 into a selection of its services. Following penetration tests and technical audits, the company aims to apply more of the model’s capabilities. For instance, the model is expected to accelerate incident response, such as when analyzing traces and reconstructing attacks.
As we often hear, humans will remain “in the loop” for security tasks. According to Orange Cyberdefense, access control, traceability, security testing of the model, and human oversight are of great importance for these applications. Critical actions remain the responsibility of experts. Feedback from real-world use will then help improve the model, or its successors, for cyber applications.
Hugues Foulon, CEO of Orange Cyberdefense, sees many benefits to using AI. For example, it helps elevate the volume, speed, and trust in security tasks to a higher level. According to him, these are the three biggest challenges for cybersecurity in the AI era. “With Mistral AI, we’re building European cyber-AI, trained by our experts and under their control,” said Foulon.
European, but not exclusive
The models run on computing power in Europe, in data centers managed by Mistral itself. Mistral’s CTO, Timothée Lacroix, refers to it as a “sovereign, reliable AI solution” for European organizations, “built and managed in Europe.”
The partnership is new but not exclusive. Orange Cyberdefense says it takes a model-agnostic approach and selects the best solution based on each client’s needs.
Read also: With Mistral Large 4, Europe is (slightly) competitive again