Malo Lecoursonnais

As part of my graduation project, I joined the DATA-AI entity and its CARS team, composed of experts in AI model security. My goal is to develop a solution to secure the use of LLMs such as GPT, Gemini, and Mistral. These language models have been widely used since the advent of ChatGPT, often in the form of agents within a RAG architecture (i.e., document base). This configuration presents vulnerabilities to attacks such as prompt injection and jailbreaking, which can lead to the extraction of confidential data. This is where dynamic fuzzing comes into play. "Fuzzing" refers to the action of generating attack prompts to deceive the LLM, while "dynamic" denotes the complexity and precision of attacks that adapt to the target AI agent. To implement dynamic fuzzing solutions, I am conducting: - State of the art of attacks: Defining and contextualizing each type of attack. - State of the art of solutions: Analyzing open-source tools like Nvidia Garak, Giskard, and Promptfoo, and identifying their pros and cons. - Evaluation: Benchmarking each solution, both quantitatively and qualitatively. - Scaling: Deploying and evaluating the behavior of solutions at different scales. - Adaptation to Orange's specifics: Tailoring solutions to internal project data, particularly LiteLLM, a proxy for LLMs. In parallel with this project, I continue my collaboration with my former team to support Orange Middle-East-Africa subsidiaries in adopting generative AI tools, in connection with our partners in chatbot publishing. I appreciate the diversity of these projects, combining technical and analytical aspects for the first, and relational and pedagogical elements for the second. I would like to thank Sébastien Durand for his trust and support on this complex new topic, as well as my mentor Catherine Billard for the flexibility she provides.

As part of Orange Innovation, I worked within a team specializing in generative AI and chatbot solutions for the Orange Group and its clients in France and the MEA region. My main responsibilities included project management, technical implementation, and client support for the deployment of innovative solutions: - Project management: Product roadmap tracking, reporting (Jira & Trello), and leading the apprentice and intern community. - Technical implementation: Development of APIs and scraping scripts (Python, Node.js), integration of containerized solutions (Docker, Kubernetes), and design of client PoCs in both French and English. - Client support: Training and support for international clients on Low Code chatbot solutions, adapting solutions to business needs. - Technology watch & innovation: Exploration of market tools and vendors, discovery of Red Hat OpenShift, and initial work on semantic vectorization. - AI & Security: Participation in projects to secure LLMs (ChatGPT, Mistral), risk analysis (prompt injection, data leakage), evaluation of red-teaming solutions, and deployment at scale within Orange's IT environment. This experience allowed me to develop cross-functional skills in project management, software development, and AI, within a collaborative and expert-driven environment.