Prompt injection detector

A web page can contain instructions that humans never see but AI assistants read when they summarize or cite it. Check a page, or paste a text before you hand it to an AI.

Check

Free, no sign-up. Pasted text is analyzed and then discarded: it is never stored.

What the detector looks for

How to read the result

• Critical: an instruction aimed at an AI sits in hidden content. This is almost never legitimate; if it is your site, check whether it has been compromised.
• High: suspicious invisible characters, or an instruction found in a text you plan to give to an AI.
• Review: an instruction-like phrase is visible on the page. Often legitimate, for example an article explaining prompt injection.
• Info: common and usually harmless, reported for completeness.

The detector uses explicit rules, not an AI model: it is fast and predictable, but it does not catch everything. A cleverly rephrased injection can slip through, and a text about injection can trigger it. For agents and RAG systems in production, combine it with defense in depth: treat external content as data, least-privilege tools, human confirmation for sensitive actions.

Developer? The detector is available as a free JSON API.

More free tools