From a security standpoint, it's better for us all for LLMs to be easily injectable. This way you can at least assume that trusting them with unvalidated input

From a security standpoint, it's better for us all for LLMs to be easily injectable. This way you can at least assume that trusting them with unvalidated input is dumb. If they are 'human level', then they will fail only in catastrophic situations, with real ATP level threat actors. Which means they would be widely trusted and used. Better fail early and often than only under real stress.