Your LLM Stack Is a 2008 Database With Better Marketing — Lovina Dmello, NVIDIA

Summary

The main theme is that production ML security breaches are rarely due to exotic AI attacks, but rather to common infrastructure mistakes. Key examples include insecure Ray clusters with open dashboards, job APIs, and authentication turned off by default. The practical takeaway is to focus on fundamental security practices like proper configuration and authentication, rather than solely on advanced attack vectors.