CodeCloak: A DRL-Based Method for Mitigating Code Leakage by LLM Code Assistants

This Black Hat conference talk explores CodeCloak, a novel deep reinforcement learning agent designed to mitigate code leakage risks when using LLM-based code assistants. Learn how this method manipulates prompts before they reach code assistant services to achieve two competing goals: minimizing sensitive code exposure while maintaining useful suggestions for developers. The presentation, delivered by Amit Finkman, demonstrates CodeCloak's effectiveness across various code repositories and its transferability between different models like StarCoder and Code Llama. Discover the methodology developed for reconstructing original codebases from prompt segments, which helps analyze leakage risks and evaluate CodeCloak's performance in real-world development scenarios.