Forget what you think you know about AI safety! As AI dives headfirst into full automation, privacy expert Katharine Jarmul warns that those shiny "guardrails" are easier to bypass than you'd expect, and models are surprisingly good at memorizing your sensitive data, creating sneaky leak risks. Turns out, just red-teaming once isn't enough, and we can't simply rely on providers to fix all our security woes.
The real talk? It's time to stop the "safety" myth. Engineers need to build a culture of iterative security, get serious with threat modeling, and maybe even consider local LLMs to truly lock things down. Your future security isn't going to magically appear with the next model update!
Watch on YouTube
Top comments (0)