What's the other 99.94% of AI security then?

I read that framework choice accounts for 0.06% of security outcomes in agentic AI. That's such a tiny sliver it makes me wonder if anyone's actually measuring the noise, or if the whole exercise is just rearranging deck chairs on the Titanic. What do you think fills the rest of that pie chart—human error, bad prompts, sloppy coding, or just luck?