Research Engineer, Mistral AI

• Hands-on expertise in offensive and/or defensive cybersecurity: vulnerability analysis, web/network/cloud security, secure coding practices, and SOC
• Strong software engineering skills: clean, reliable, well-tested code — not just exploit scripts
• Fluency in Python plus comfort reading lower-level languages (C/C++) for vulnerability analysis
• DevOps experience: Docker, Kubernetes, cloud deployments, sandboxed or simulated environments
• A pragmatic research-to-engineering mindset: ship a working first version, then refine and scale it
• Working knowledge of RL techniques and LLM training methodologies, or strong motivation to develop it
• Self-starter, low-ego, collaborative — comfortable working across research and engineering

Desirable

• Cybersecurity is one of the areas where AI could do the most good — and one where the capability must be built with particular care.
• As a Research Engineer on this track, you'll push our models' capabilities in vulnerability discovery and remediation, security analysis, incident response, and across the offensive-to-defensive spectrum of cybersecurity, and it will be your job to develop that ability deliberately and safely
• The work sits at the meeting point of research and engineering: you'll devise new approaches and be the one to implement them.
• Concretely, that means designing and building RL environments that reproduce realistic security scenarios, pipelines that generate synthetic training instances at scale, and experiments and evaluations that show what our models can actually do.
• Your results feed directly into the production training runs that shape our frontier models, in close collaboration with Mistral's researchers, engineers, and security specialists
• Security specialists and ML engineers are both welcome here — what matters is that you can hold your own in both worlds, and are eager to go deeper on the one you know less
• Design and implement RL environments simulating cybersecurity scenarios
• Build the pipelines and infrastructure that generate synthetic cybersecurity training instances at scale — thousands of scenarios, not one hand-crafted exercise
• Conduct experiments and evaluations of model capabilities on these environments, from quick prototypes to controlled benchmark runs
• Own the infrastructure behind the environments: containers, sandboxes, VMs, cloud deployments, and orchestration (Kubernetes), all managed as code
• Partner with researchers and security specialists across Mistral, distilling their domain expertise into reproducible environments and datasets
• Write clear, efficient code in Python and enforce strong software-design practices: testing, code review, CI/CD