Forget what you think you know about AI safety! Privacy expert Katharine Jarmul spills the tea on why those fancy AI "guardrails" are basically paper-thin and easily tricked. As AI takes over more tasks, engineers can't just cross their fingers and hope model providers magically fix privacy — it's on them.
Turns out, these smart models often "memorize" your sensitive data, not just learn from it, and common attacks can bypass filters with ease. The talk pushes for better, iterative security, smart threat modeling, and even suggests using local LLMs to beef up defenses. Stop one-and-done red-teaming and start building a real risk radar!
Watch on YouTube
Top comments (0)