Prompt Injection Detector

Detects prompt injection: text that tries to take control of the AI assistant reading it. The text comes with its source, because that changes the answer: a user may give the assistant instructions, but a retrieved page, email or tool result may not. Jev answers one Noul and returns P(injection); code blocks at a threshold you tune with the sweep. Against it: the phrase deny-list most filters start with ($0), and an LLM classifier, which has to read the very text that may be attacking it. Measured: accuracy, injections missed, safe inputs blocked, and the threshold sweep.