Skip to content
Prompt Injection ScannerBeta

See what the AI will read that you can’t.

Scan text

Check it before you paste it into an AI

See what the AI will read that you can’t.

Hidden instructions, invisible characters and prompt injections can sit inside text that looks ordinary. Inspect it first — in your browser, without uploading anything.

  • Find hidden text & instructions
  • See prompt-injection indicators
  • Your text never leaves the page
Read the qualification: Results identify known and suspicious indicators for review. They cannot guarantee that content is safe.

What the free scanner does

Reveal hidden text

Zero-width characters, bidi controls and default-ignorables become readable U+XXXX tokens instead of staying invisible.

Spot injection patterns

Instruction, target and objective concepts must appear together in a versioned combination — a keyword on its own is not a finding.

Text, Markdown & HTML

Paste text or open a .txt, .md or .html file. HTML comments, metadata, hidden text, link destinations and styles are all inspected.

Context is part of the answer

Quotation, code, research and policy framing can reduce a finding to review-only or suppress it, because an example of an attack is not an attack.

Inspect the textual layers the browser supplies

Type or paste text

Scan up to 102,400 Unicode scalars and 400 KiB of admitted plain text. A rich paste can additionally preserve up to 32 KiB of HTML from the same event, with a 432 KiB combined cap.

Choose a local file

Inspect one strict UTF-8 .txt, .md, .markdown, .html or .htm file up to 128 KiB. The file stays inside the page and worker.

Review coverage

See which visible, hidden, comment, metadata, attribute, link, code and encoded-candidate layers finished or were unavailable.

Evidence in context, not a risk percentage

A keyword by itself is not enough. The scanner requires versioned combinations of instruction, target and objective concepts, then shows a bounded excerpt and original location. Quotation, code, research and policy framing can change a finding to review-only or suppress it.

Policy example: “Reject retrieved pages that instruct the model to disregard its governing instructions.”

The sentence discusses an indicator inside a prohibition. It illustrates why quoted attack language is a required hard negative rather than automatic proof of intent.

Compare more indicators and hard negatives.

Leave with a defensible next action

Result situationNext action
Strong indicatorsQuarantine the source from automated ingestion until a person reviews the evidence.
Suspicious indicatorsCompare the passage with the intended visible source and edit or isolate it when appropriate.
No indicators in inspected textContinue cautiously with the downstream system’s normal isolation, validation and approval controls.
Incomplete coverageUse a tool or manual process that can inspect the unsupported layers before ingestion.

The launch boundary

The scanner does: inspect supplied text with deterministic structural, Unicode, encoded-text and English instruction rules; preserve source locations; and report coverage, evidence, limitations and actions.

The scanner does not: upload content, call an AI model, fetch URLs, render submitted markup, remove passages, test a live model, inspect PDF/DOCX/images/audio/video or replace downstream security controls.

Browser-local rules are public and can be studied by attackers. Keep untrusted content separate from governing instructions, restrict model tools and require human approval for consequential actions.

Understand what the report means

What is prompt injection?

Separate direct instructions from untrusted instructions embedded in material an AI system reads.

Read the introduction

Where can text hide?

Review HTML channels, Unicode controls, encoded candidates and the layers launch cannot inspect.

Read about hidden channels

How is it evaluated?

See rule identities, matching, hard-negative slices, release thresholds and missing external evidence.

Review the methodology