Design Patterns for Securing LLM Agents Against Prompt Injections

I saw Invariant Labs and I upvoted this. :) Definitely reading it as well.

Looks promising, would this be possible to create a whole new language?