LangChain Community,
Hello. I am recently reading about new and emerging hardware-based solutions for AI safety and security [1].
With respect to LangGraph agentic workflows, while sandboxed AI agents might be granted sets of permissions [2], altogether, only some of these might be needed at individual workflow nodes.
Brainstorming, security middleware could be a way to go to encapsulate hardware-interoperability features. Node and/or state metadata might also be useful for describing concrete permissions which can be known a priori or, perhaps, to provide functions which return concrete permission object instances as a result of processing configuration-, context-, and/or state-related data.
How do you see developer convenience with respect to using new and emerging hardware-based solutions for AI safety and security?
[1] NVIDIA Open Agent Safety Platform
[2] Sandbox Policies | NVIDIA OpenShell