Every check we run, and every one we don’t
27 detectors and five models, all running in your browser. Below is the whole list — what each one finds, how sure it can be, and the things it is known to miss. You should not have to take a privacy tool’s word for any of it.
- 27 detectors, all switched on
- Rule pack 2026.10.6
- Nothing is uploaded
Why it matters
Three places a file goes, and what goes with it
None of this is visible when you look at the file. All of it travels with it.
Posting to social media
You think you are sharing: You are sharing a picture of your dog in the garden.
What actually goes: The file also carries the GPS coordinates the camera recorded, to within a few meters — your garden. Plus the faces of anyone in shot, the house number behind them, and the plate on the car in the drive.
Some platforms strip location on upload. Some do not, and none of them strip the things that are visible in the picture itself.
Uploading to ChatGPT, Claude or another AI tool
You think you are sharing: You are asking a chatbot to read a screenshot or a document for you.
What actually goes: Whatever is in the frame goes with it — a colleague’s email address, a customer list, an API key in a terminal window, the patient name at the top of a form.
Once it has been sent it has been sent. Check what is in the picture before it leaves, not after.
Forwarding a Word document
You think you are sharing: You are passing on a document somebody sent you.
What actually goes: Word files keep tracked changes, review comments, the author’s name, and text that was deleted before it reached you. A black rectangle drawn over a paragraph usually leaves the paragraph underneath it, selectable.
This is the most common accidental leak in documents sent outside an organisation, and the sender almost never knows it happened.
Hidden file data
What your camera wrote down without telling you
Read from the original bytes, before anything is resized. That ordering matters: a downscaled copy has usually already lost its EXIF, so a tool that resizes first finds nothing and reports nothing.
| What | Risk | Why it matters | Read from |
|---|---|---|---|
| GPS location | High | Where the photo was taken, usually to within a few meters. Posted publicly, that can be your home, your workplace or a child’s school. | EXIF.GPSLatitude |
| Document author | Medium | The file properties name whoever wrote or last edited it — often someone who never expected to be named outside the company. | core.creator |
| Camera or phone model | Low | Make, model and sometimes a serial number. It links separate photos back to the same device, and so to you. | EXIF.Model |
| Date and time taken | Low | Recorded to the second, which can place you somewhere at a particular moment. | EXIF.DateTimeOriginal |
| Software that made the file | Low | Names the app and version. Mostly harmless alone, but it tells an attacker which software to target. | EXIF.Software |
Every one of these is stripped from the copy you download, and you are told which ones were there. The image data itself is copied through untouched — stripping metadata costs no quality.
In the picture
Five models, all of them running on your device
Named rather than described. “AI-powered detection” tells you nothing; the model, the way it is run and what it is known to miss can be checked.
MediaPipe BlazeFace
Faces, including small ones in the background
Full-range model, run over the whole picture and again over a 3×3 grid of overlapping crops. The tiling is what catches a face in a crowd — without it the model shrinks the whole image to a small square first and distant faces vanish.
Tesseract OCR
Any readable text in the picture
Everything it reads is handed to the twenty-seven detectors below. English only today, which is a real limit rather than a detail — see the honest list at the bottom.
Plate localiser + second OCR pass
License plates on vehicles
Finds the plate by its shape — a bright, dense rectangle of characters — crops it, enlarges it, and reads it with a recognizer set up for one line of capitals and digits. A plate is only reported when there is a vehicle under it.
zxing-wasm
QR codes, barcodes and boarding passes
Decoded from the original full-resolution bytes, where a dense code is still legible. A boarding pass barcode carries your name and booking reference in plain text.
EfficientDet-Lite0
ProWhole people and vehicles
Covers a person head to foot rather than just their face — useful when a uniform, a build or a pushchair identifies someone as readily as their face does.
Detectors
What runs over the text, and how sure it can be
This split is the most useful thing on the page. A card number that passes a Luhn check is a fact. A street address is a judgement. Presenting both as one undifferentiated list of “detections” is how tools train people to ignore their warnings.
Checked by maths — these are not guesses
Each of these has a checksum or a format rule built into the standard, so a match can be verified rather than estimated. A false positive here is close to impossible.
- Credit and debit cards— Luhn checksum
- IBANs— ISO 13616 checksum and per-country length
- Passport MRZ— Check digits on the machine-readable zone
- US Social Security numbers— Assignable-range rules
Matched by pattern
Strong, well-defined shapes. These are reliable in practice, and each one is validated beyond "it looks about right".
- Email addresses
- Phone numbers— International and local formats
- API keys and secret tokens— Known provider prefixes
- IP addresses— IPv4 and IPv6, private ranges excluded
- Boarding passes— IATA BCBP barcode payloads
Judgement calls — we tell you when it is one
Things with no checksum to lean on. We show you the box and let you decide, rather than quietly deciding for you.
- License plates— Only when a vehicle is under them
- Street addresses— Suffix and format heuristics, UK postcodes, US city/state/ZIP lines
- Dates of birth— Plausible-date rules
- Faces— Model confidence, adjustable
- Signature blocks— DocuSign, Adobe Sign and other e-signing stamps, with the name under them
- Personal names— Anchored to honorifics, labels like “signed by”, or a list of common given names — a judgement, so you get the box, not a verdict
Inside Word documents
A different pass entirely. None of this is visible when you open the file, and all of it travels with it.
- Tracked changes— Text deleted before it reached you
- Review comments— Invisible in print, visible to anyone who opens it
- Hidden content— Hidden text, rows and columns
- Embedded files— Another file travelling inside this one
- Fake redactions— A black box drawn over text that is still underneath it
- Author names— In the file properties
What you control
It opens covered. You decide what gets uncovered.
That ordering is the whole design. The alternative is handing somebody a photograph and asking them to spot what they forgot — which is the thing that failed in the first place.
Decide what is covered
- Everything we found opens already covered — you uncover what you want seen, rather than hunting for what you forgot.
- Tap a category to uncover or re-cover that whole group at once.
- Click any single box to change just that one.
- Drag across the picture to cover something we missed. Click it again to remove it.
- Undo, Redo, and “Back to auto” to return to what we found.
Decide how it looks
- Blur, pixelate or a solid box — set per category, not all-or-nothing.
- Pixelate at full strength is the default, because it is destructive rather than cosmetic.
- A strength slider, and black or cream for solid boxes.
- “Keep the main person visible” — covers the crowd behind your subject and leaves your subject alone.
Check our work
- A compare slider: drag between the original and the covered version.
- Outlines on or off, so you can see every region we found including the ones you uncovered.
- A full list of findings with masked excerpts and the actual tag names, so you can tell a real detection from a wrong one.
- On export: the covering is painted into the pixels, every hidden field is stripped, and you are told exactly what the copy no longer carries.
Honest limits
What we don’t catch
On a tool like this, the list of things it cannot do is worth more than the list of things it can. You will find out which kind we are the first time something gets through — better to read it here.
Text recognition is English only
Text in other scripts is not read, so nothing is found in it. If your document is in another language, treat the result as "we checked the hidden fields" and read the picture yourself.
License plates need a visible vehicle
No car in the picture means no plate search. That is deliberate: it is what stops us drawing a black box over plate-shaped graffiti on a wall and calling it somebody’s registration.
Very small pictures skip the models
Below 64 pixels there is no readable text and no recognisable face, so we do not download twenty megabytes of model to prove it. The hidden-field check still runs — a thumbnail can still carry GPS.
Hidden worksheets are reported, never deleted
Removing one can turn a working spreadsheet into a page of #REF! errors. We tell you it is there and leave the decision to you.
Detection is best-effort
We will never tell you a file is "safe" or "clean". The most we say is "no issues found", because nothing matching is not the same as nothing being there. Look over anything before you share it.
All of it runs in your browser. Nothing is uploaded.
The models are downloaded to your device once and the picture never leaves it. The only request the page makes is one anonymous line: the kind of file, which categories were found, and the risk level. No filename, no content, no previews.