Speaker
Description
AI coding agents are becoming part of everyday scientific work. Yet every file, prompt and test output an agent reads may leave the institute. This puts intellectual property at risk, both for new research results that we may want to patent and in joint academic–industry projects. Moreover, contracts or data protection law may forbid passing sensitive data to third parties altogether. As AI ambassador of HUN-REN Wigner RCP, I present practices that keep the benefits of AI while limiting this exposure. We run agents in isolated environments that work on both Linux and Windows. We explore how to set up devcontainers, which separate the agent from the rest of the machine and keep builds reproducible, from C++ to GPU-enabled PyTorch. We review how to work on projects that need secrets, such as access tokens and API keys. We show how to keep these secrets out of agent transcripts. We set up a VS Code environment that uses the GenAI4Science models of HUN-REN Science Cloud, so model requests stay on HUN-REN's own servers. We also point out less obvious leaks and show how the choice of agent, model and account decides whose data terms apply. The presentation includes a demo that shows the setup step by step.