How to use Hidden Text X-Ray
- Choose your source. Paste text or HTML, or select a file from your device. Local files are read in your browser.
- Add pixel reading if needed. In the file tab, open Read text inside pictures, PDFs and videos, then press Read words in this file for optional English text reading. Choose PDF page and video sampling limits first. Words inside pictures are read only when you choose the optional reading button.
- Press X-Ray it for basic inspection. Read the summary, then compare the approximate visible view with the extracted computer-readable text.
- Open the evidence. Each finding shows the exact text, its location, the hiding method and why it may matter. Encoded messages have a separate decoded view.
- Check the limitations. A partial inspection is not a complete picture. For web pages, use View page source, then paste that source.
- Copy or export if needed. You can review and copy text for pasting into AI or download a text or JSON report. Clear removes the current input and results from this page.
If you enjoy what we do, consider supporting us. Every little bit means the world! Support Catpool on Ko-fi
What it checks
Hidden text is not always harmful. Accessibility labels, document properties and HTML comments are everyday parts of files. The useful question is what they say and whether they belong there.
- Invisible characters: zero-width marks, bidirectional controls, unusual spaces, Unicode tag messages and readable variation-selector byte encoding.
- Disguised words: Latin mixed with Greek or Cyrillic letters, with common look-alikes mapped back to Latin.
- HTML hiding: inline and simple local CSS, hidden attributes, tiny or low-contrast text, off-screen markers, templates, comments, metadata and alternative or attribute text.
- Encoded text: readable base64, hex, percent-encoding and runs of numeric HTML entities.
- Office files: direct hidden, white or tiny text; stored revisions, comments, properties, spreadsheet hiding markers and presentation notes.
- PDFs and pictures: basic PDF hiding markers and image metadata, plus optional PDF.js text extraction and English OCR of selected rendered pages, images and video frames. Low-contrast and tiny-text regions are approximate hints.
- AI-aimed instructions: weighted language patterns for instruction overrides, secrecy, commands, credentials and possible data exfiltration. Concealed matches receive more weight.
What it cannot check
This is not a virus scanner, forensic suite or AI authorship detector. Optional English OCR can attempt printed text in images, selected scanned-PDF pages and sampled video frames. It can miss or misread faint, small, stylised, handwritten or non-English text. Unsampled pages, frames and audio are not checked. It does not run scripts, load external stylesheets, sign in to pages, open attachments or reproduce every feature of Word, Excel, PowerPoint or PDF viewers.
Encrypted files, complex PDF font mappings and compressed object streams may not yield text. Office formatting inherited from styles and themes can be missed. A browser needs DecompressionStream support for compressed file parts. Inputs have size and processing limits, which are reported when reached.
Nothing found by these checks does not mean a file is safe. A finding is not proof of malicious intent either. An AI instruction can be a harmless quotation, and legitimate typography can use invisible characters.
Privacy
Paste and file inspection runs locally using browser JavaScript and a local worker. The tool does not upload, log or save your text or files, and does not follow links or run content scripts. Optional Tesseract OCR and PDF.js run in your browser. Their engine, English model and PDF support files are served from Catpool, not a third-party CDN, and may be retained by normal browser caching. The first OCR engine/model load is about 6 to 10 MB; PDF rendering may load several more MB. These assets are reused within the page session, while later visits depend on browser caching. Recognised text and image pixels are not sent to an AI service. The same-origin reading assets are supplied by Catpool.
Web address checking is not connected. No request is made.
Downloaded reports contain the findings you inspected, potentially including private metadata. Keep them secure. Clear resets this tool, but cannot erase your clipboard or downloaded reports.
FAQ
What does the OCR confidence number mean?
It is the recognition engine’s own score, not a probability that the words are correct. Check the local pixel preview and exact location. Faint, enhanced, stylised and handwritten text can be misread even with a high score.
What do red boxes and contrast enhancement show?
Red boxes mark approximate recognised regions with low sampled contrast, small height or text recognised only after enhancement. The Original and Contrast enhanced buttons compare local pixel views. Enhancement can introduce artefacts, so it does not prove deliberate hiding.
Does it check the whole video?
No. It reads sampled frames at your chosen interval, up to your frame limit, without uploading the video. Audio and unsampled frames are not checked. Your browser must be able to decode the file’s codec.
Do the bit statistics prove there is a secret message?
No. Ordinary photos, resizing, editing and compression can produce similar least-significant-bit statistics. They are information-only hints. This tool does not recover or rule out arbitrary steganographic messages.
Is this different from Text X-Ray?
Yes. Text X-Ray looks for scam patterns in messages. Hidden Text X-Ray looks inside text and supported file structures, showing concealed content and possible AI instructions.
Can it tell whether text was written by AI?
No. It looks for hidden content and instructions aimed at AI, not who wrote something. It does not make an AI authorship claim.
Does an AI instruction mean someone attacked me?
No. Documentation, examples and quoted discussions can contain the same phrases. Look at the exact wording, where it was stored and whether it makes sense in context.
Will it find every hidden message in a file?
No. It inspects supported text and metadata structures, not every possible format or hiding technique. Optional OCR attempts printed text in pixels, but is not complete. Bit statistics and trailing data are only hints, not proof of steganography. The report states the limits of that inspection.
Are my files uploaded?
No. Selecting or dropping a supported file reads it locally in your browser. The tool does not upload the file. Optional OCR loads its engine and English model from Catpool, then recognises pixels locally. Exporting creates a report on your device. Normal browser caches can retain the engine assets, not the tool’s input.
Why are accessibility labels and comments listed?
They can contain text that is absent from the normal visual view. Most are useful and harmless, so they are usually marked as information, not a high-severity warning.
Is the safe copy a repaired or guaranteed-safe document?
No. It is a text-only approximation with invisible controls removed, not a repaired file. It may change legitimate scripts or emoji and may still contain visible instructions. Review it before using it.
What should I do with a suspicious link?
Do not follow it from a finding. Use Link X-Ray for address checks, or open the organisation’s known website yourself. This tool never makes extracted links clickable.
