Loading page content
Scans input for eighteen known prompt-injection patterns (instruction overrides, role manipulation, delimiter escapes, encoded payloads) and returns a 0 to 100 risk score with four-tier severity per finding.
Scan prompts and user inputs for common prompt injection patterns. Detects instruction overrides, role hijacking, control token injection, jailbreak attempts, and more. All client-side, nothing sent to a server.
Scans for 18 known prompt injection patterns including DAN/jailbreak, instruction override, control token injection, role hijacking, and encoded payloads. Client-side heuristic analysis. Use alongside server-side protections.
Prompt injection is not a solved problem and this is not a solution to it. Pattern matching catches known phrasings (ignore previous instructions and its many variants) and misses novel ones, encoded payloads, and anything expressed in a language the patterns were not written for. Treat a clean scan as no evidence of a problem, never as proof of safety.
The structural defenses are what actually matter: never let model output trigger a privileged action without a separate authorization check, keep untrusted content clearly delimited from instructions, and give the model the narrowest tool permissions the task allows. Indirect injection (a payload sitting in a document or web page the model retrieves) is the harder variant and the one that has produced real-world incidents.