Hold up! Thinking AI is safe because of current guardrails? Think again. Privacy expert Katharine Jarmul busts myths, revealing how easily AI models, especially LLMs, can be tricked or leak sensitive data they've "memorized." Don't count on model providers to sort out your privacy woes; engineers need to step up.
Forget one-time red-teaming fixes! It's all about iterative security, understanding architectural realities like overparameterization, and embracing solutions like local LLMs and diverse providers. Building a culture where security is paramount is key to catching those sneaky incidents before they blow up.
Watch on YouTube
Top comments (0)