Prompt Injection Detector (Heuristic scan for jailbreaks)

What is Prompt Injection Detector?

Prompt injection is the top security risk for LLM applications: untrusted text (a user message, a scraped web page, a document) contains instructions that hijack the model — 'ignore all previous instructions', 'you are now DAN', 'print your system prompt'. This tool runs a set of heuristic pattern rules over input to flag those well-known attack shapes and produce a weighted risk score, useful as a fast pre-filter and for teaching what injection looks like. It is explicitly not a complete defense: robust protection requires isolating untrusted input from instructions, filtering outputs, and giving any tools the model can call least privilege. Treat a low score as 'no obvious attack', not 'safe'.

How to scan for prompt injection

1. 1Paste the untrusted input that will be inserted into your prompt.
2. 2Review the risk score and each flagged pattern with its severity.
3. 3Use it as one signal alongside input isolation and output filtering.

Key Concepts

Prompt injection

Untrusted text that smuggles instructions to override the model's intended behavior.

Jailbreak

Input that tries to bypass safety guardrails, often via role-play personas.

Tips & Best Practices

• ▸Never concatenate untrusted input directly with your system instructions — keep them in separate message roles.
• ▸A low risk score means no known pattern matched, not that the input is safe. Novel attacks evade pattern matching.
• ▸Pair detection with output filtering and least-privilege tool access for real defense.