Prompt Injection Scanner
Paste a document, email, web page or tool output before an AI agent reads it. The scanner looks for hidden instructions: invisible Unicode, hidden HTML, instruction-like phrases, encoded payloads and data-leaking image links, then shows exactly where each one is and gives you a cleaned copy.
Everything is checked locally. The scan looks for the techniques used in real attacks on AI agents: invisible “tag” characters that spell hidden messages, zero-width encodings, text hidden with CSS, instructions in comments and image alt text, base64 payloads and image URLs that leak data.
| Severity | Type | What we found | Where | Text |
|---|
The cleaned text drops invisible characters, bidirectional controls and look-alike letters, and (for HTML) comments and hidden blocks. The safe wrapper tells the model to treat the text as data. Neither makes untrusted text harmless: also limit what the agent can do and require approval for actions.
What is prompt injection, and what does this scan for?
Prompt injection is when text an AI reads, such as a web page, email, resume or document, contains instructions meant to take over the AI. Indirect injection hides those instructions in content the user never wrote, so an agent that summarizes your inbox or browses the web can be steered by a stranger. This scanner checks for the techniques used most often.
Invisible Unicode
Unicode “tag” characters (U+E0000 range) display as nothing but many models read them as ordinary letters, so a message can ride along inside normal-looking text (“ASCII smuggling”). The same trick works with variation selectors and runs of zero-width characters. Bidirectional controls can make text read differently than it is processed, and look-alike Cyrillic or Greek letters inside Latin words evade keyword filters.
Hidden HTML and Markdown
Text set to display:none, zero font size, white-on-white or moved off-screen, HTML comments, and instructions in alt, title or aria-label attributes are invisible to readers but visible to an AI that receives the raw HTML.
Instruction-like phrases and fake markers
Phrases such as “ignore previous instructions”, “do not tell the user”, “if you are an AI assistant”, requests to reveal the system prompt or send data to a URL, and fake chat-template markers (<|im_start|>, [INST], System:).
Encoded payloads and data leaks
Base64, hex or percent-encoded blobs that decode to instructions, ROT13 text, and Markdown or HTML images whose URL carries placeholders or long data in the query string. Rendering such an image sends conversation data to an attacker’s server.
Limits, and how to actually defend an agent
A pattern scanner catches common attacks, not every attack, and a clean result is not a guarantee. Treat all outside content as untrusted data, give agents the least privilege they need, require human approval for sending, deleting or paying, never let one agent session mix untrusted content with private data and an outbound channel, and log what agents do. See the Prompt Injection Reviewer skill and the Hidden Characters detector.
How to use the Prompt Injection Scanner
- Paste the text, HTML or Markdown an AI agent is about to read, or open a file.
- Read the verdict and the findings: each shows its line, column and the text involved.
- Check any decoded hidden messages, then copy the cleaned text or the safe wrapper.
- Download the report for a security review.
Why use this scanner
- Finds invisible-Unicode, hidden-HTML and encoded attacks
- Decodes hidden messages so you can read them
- Shows exactly where each finding is
- Private: nothing leaves your browser
Related tools
Frequently Asked Questions
What is prompt injection?
Prompt injection is text that tries to give instructions to an AI model that the user never intended. Indirect prompt injection hides those instructions in content the AI reads on the user's behalf, such as a web page, email, resume or PDF, so a stranger can steer an agent that summarizes, searches or takes actions.
Does a clean result mean the text is safe?
No. The scanner recognises common attack patterns and encodings, but new techniques appear constantly and natural-language attacks can be phrased in endless ways. A clean result lowers risk; it does not remove it. Always treat outside text as untrusted data and limit what the agent can do.
What are invisible Unicode tag characters?
Tag characters (U+E0000 to U+E007F) are a Unicode block that renders as nothing in most apps, yet many language models read them as the matching ASCII letters. An attacker can hide a full instruction inside an ordinary-looking sentence. This tool decodes such runs so you can read the hidden message.
Is my text uploaded anywhere?
No. The scan runs entirely in your browser with JavaScript. Nothing you paste or open is sent to a server.
What does the cleaned text remove?
It removes invisible tag characters, bidirectional controls, zero-width and filler characters and soft hyphens, maps look-alike Cyrillic and Greek letters back to Latin inside Latin words, and for HTML also removes comments and elements hidden with CSS or the hidden attribute. It does not remove instruction-like sentences that are visible; review those yourself.
How do I protect an AI agent from prompt injection?
Treat all external content as data, not instructions; give the agent least-privilege tools; require human approval for sending, deleting, paying or sharing; keep untrusted content, private data and outbound channels out of the same session; clearly delimit untrusted text in the prompt; and log agent actions. Scanning is one layer, not a complete defense.