Build your own prompt injection classifier with decision models, or enable the check directly

Opper AI's blog describes how to build your own prompt injection classifier with decision models, and compares four decision models against one LLM on 6,556 prompts from public datasets. The decision models return a thresholdable probability in about 200 milliseconds, roughly ten times faster than the LLM, with similar quality; the same GPT-6 Luna took 271 milliseconds as a decision model and 1.7 seconds as an LLM answering. All models missed many injections hidden in tool results, with more than a quarter not caught. Opper's built-in Prompt injection check asks about tool results separately and calibrates thresholds by model and input, catching 94% of injections hidden in tool results in real calls, with about half the false positives, adding about 200 milliseconds itself. The author says roleplay and fictional framing are the most common missed types, and non-English attacks have the lowest detection rates.