OFFLINE HEURISTIC
Prompt Injection Scanner
Paste a prompt, email, page, or file. Matching stays in this browser — nothing is sent anywhere.
0 chars · Ctrl+Enter
No scan yet.
What gets flagged
Each category has a weight. The score is their sum, capped at 100. 0–27 is Low, 28–59 is Medium, 60–100 is High. One textbook override rates High. A lone “act as” stays Low.
- 62Instruction overridePhrases that tell a model to drop, forget, or replace earlier instructions. This is the classic direct injection.
- 18Role hijackingAttempts to redefine who the assistant is. A lone “act as” is often a normal prompt; it matters more when stacked with other hits.
- 44System prompt extractionAsks the model to reveal hidden instructions, the system prompt, or text that came before the user message.
- 64Delimiter and token injectionFake chat-template tags (ChatML, Llama, Mistral, and similar) used to spoof a system or assistant turn.
- 36Encoding smuggleBase64, rot13, hex, or similar encoding next to instruction language, or a blob that decodes into injection text.
- 48Jailbreak and unrestricted modeDeveloper mode, DAN, jailbreak, or language about disabling safety filters and guardrails.
- 30Indirect model addressInstructions buried in ordinary content that speak to the model directly — common in poisoned pages, emails, and files.