What 78K attack samples taught me about catching prompt injection

I spent the last while building a prompt-injection detector trained on 78,000+ attack samples. Here's...