Comment by datadrivenangel

The whole point of "agentic AI" is that you don't have to rigorously test every potential interaction, which means that even a temperature zero model may behave unexpectedly, which is bad for security.