Comment by fxtentacle

That paper says you need to control "0.1% of the training data size" for a 40% chance for one single injected prompt to fire. So that's millions of images or billions of text tokens for real-world models.