Is this prompt an injection?

Paste what a user typed. A guard habit trained on 300 labelled prompts and the model alone answer the same question side by side. On held-out prompts the habit had 0.8% false alarms against the model's 39.8% (benchmark).

How the habit gets better

You send the prompts your app sees with a yes/no label, for example from flags your moderators already raise. The habit retrains on them and answers on its own once it beats the model on held-out labels. Anything it is unsure of goes to the model. Fifty labels took false alarms on new traffic from 15.4% to 4.1%.

Built for user-typed prompts. On tool output and documents the model alone does better; the benchmark shows where. Examples on this page come from jackhhao/jailbreak-classification and may include prompts the habit trained on; the benchmark scores only held-out ones.