This language already does that by "ignore previous instructions" as function call and "what happened at tiananmen square?" as function return.
The solution is obvious. AI can write a transpiler from this unsafe language to safe language which have 1:1 mapping to unsafe language, working on safe language, then transpile back to unsafe language.
I could come up making a transpiler before even finish reading this article. No reason the current AI models can't do it.
The solution is obvious. AI can write a transpiler from this unsafe language to safe language which have 1:1 mapping to unsafe language, working on safe language, then transpile back to unsafe language.
I could come up making a transpiler before even finish reading this article. No reason the current AI models can't do it.