There are quite a few hard problems in securing AI systems. We need to answer difficult questions just to model the threats, let alone to come up with solutions. What are the real world assets in an agentic AI system? How do we define an objectively “bad” outcome, anyway? Still, AI is here and we need to cope with what we’ve got now, even if it’s by using imperfect shortcuts. One area where I believe shortcuts can be made relatively effective is Physical AI. This area, consisting of AI used in robots, vehicles, industrial systems, etc., may or may not be easier to model, I don’t know, but it does allow for easier specification of guardrails, which may help address the overall Physical AI security problem.

Let us discuss what guardrails are, and how they can be implemented in Physical AI systems to help address AI security challenges.