I am now surprised that LLMs don't already have a way to clearly identify prompts coming from their user. Would be great to have a secure, priveleged path from my keyboard to my local model / model harness, and all other prompts disregarded. But, then that would totally break agentic workflows because the agent needs to read and act upon information it gets from the harness and environment in order to operate agentically.
I suspect a solution to this problem is one of those things that in a couple years will be baked in to everything.
Schlagbohrer•4m ago
I suspect a solution to this problem is one of those things that in a couple years will be baked in to everything.