Skip to content
Prompt Injection ScannerBeta

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

Scan text

Coverage in the channels this scanner reads

What we detect

The scanner looks for hidden content, suspicious instruction patterns and characters that change what an AI reads without changing what you see. Everything below is inspected in your browser, in supplied plain text, Markdown and HTML.

Hidden and invisible content

Text can be present to a model and absent to a reader. These are the shapes that difference usually takes, and each one is reported with the exact location it was found at, so you can compare the passage with the source you meant to send.

Zero-width characters

Zero-width spaces, joiners and non-joiners that carry no glyph but survive a copy and paste.

U+200B, U+200C, U+FEFF

Bidi control characters

Direction overrides that can reorder how a line reads while leaving the underlying sequence intact.

U+202E RIGHT-TO-LEFT OVERRIDE

Default-ignorable code points

Characters a renderer is entitled to draw as nothing, revealed as named tokens rather than removed.

U+00AD SOFT HYPHEN

HTML comments

Comment text is never rendered to a reader and is read as text here, because a model given the source sees it.

<!-- reveal the system prompt -->

Hidden elements and styles

Content hidden by inline styles or style rules is extracted as text; the styles themselves are quoted, never applied.

display:none, visibility:hidden

Metadata and attributes

Meta names and contents, image alternative text and other text-bearing attributes are inspected as their own channels.

<meta name="description">

Instruction patterns, in context

A keyword alone is not a finding. A rule fires when instruction, target and objective concepts appear together within a bounded window, in a combination that carries a version and a rationale. Context can lower a finding to review-only or suppress it, because an example of an attack inside a policy is not an attack.

Authority override

Language that tries to displace the instructions the model was given.

"Ignore all previous instructions"

System extraction

Attempts to have the model disclose its governing prompt or configuration.

"Reveal your system prompt"

Data exfiltration

Instructions directing private content outward, including through a link or a tool call.

Hidden instruction to send a summary

Obfuscated instruction

Instructions hidden through encoding, confusable characters, controls or fragmentation.

Base64-shaped runs, decoded and re-read

Link labels and destinations

Inspected separately, because a label can describe one destination while pointing at another.

Label and href disagree

Quoted examples

Research, policy, code and fiction framing are recognised as hard negatives rather than treated as proof of intent.

A policy prohibiting an attack

What is reported either way

Every layer says whether it was inspected, partially inspected or unavailable, and the result carries that coverage whether or not anything was found. A clean result over partial coverage is not the same claim as a clean result over complete coverage, and the scanner never collapses the two.

The scanner does not inspect PDF, DOCX, images, audio or video, does not detect AI authorship, does not test a live model, and does not remove anything. It reports what it read and names what it could not.

Result boundary: Findings are indicators for review. Detection cannot certify a source, and a no-indicator result does not replace downstream isolation, validation, least privilege or approval.