That is one risk. Humans at the other end of the screen are effectors; nobody is worried about AI labs piping inference output into /dev/null.