← Secure tools

Guide

AI upload privacy and secret detection

Pasting a log, a config, or a customer note into an AI chat copies that text to someone else’s service. A local pattern scan can catch common secrets before that copy happens. It cannot decide that the rest of the text is harmless.

What leaves the device, and what does not

This site’s AI upload page runs its rules in a same-origin worker. It does not call an AI API. The moment you later paste the same text into a hosted chat, that vendor’s terms apply and the text leaves the browser. Scanning here is a pause, not a tunnel.

The page accepts pasted text and text files within 50 MB. Photos and PDFs are out of scope; scan those with their own tools, and scan any text you copy out of them.

Which patterns are in scope

The rules look for common shapes: email addresses, US-style phone numbers, cloud access-key prefixes, API-key prefixes such as the well-known sk- form, database URLs with passwords, and a few other credential markers. Previews in the result are masked. A database URL should not show the password in full. An access key should not show every character.

A paragraph about a product release, a version number, and an office extension can produce zero findings. That is expected. A 10-digit string can match the phone rule even when the separators are missing. That is also expected, and it is why you still read the matches.

False confidence and missed secrets

People, street addresses, internal hostnames without a credential shape, and one-off passwords that look like ordinary words are easy to miss. Secrets split across lines, or encoded, can also miss. A clear result is not permission to upload a customer list or a medical note. Those documents are sensitive even when no regex fires.

The other failure is noise. A public sample key from a vendor’s documentation can look like a live key. The AWS documentation key prefix AKIA is a famous example of a string that should never be a production credential, yet a scanner may still flag that shape. Investigate the match; do not assume every hit is an incident.

From a finding to a safer paste

The upload scanner does not edit the text for you. If you want replacements, open the text redactor, keep only the matches you intend to change, and read the verification line. Then paste the edited text, not the original. Clear the scanner when you are done so the textarea is not left on a shared screen. Clearing the page does not revoke a copy you already sent to an AI vendor.

A fictional example

A note says: email [email protected], phone (415) 555-0198, and a postgres URL whose user is admin. The scanner should list the email, the phone, and the database URL, with the password masked. A second note that only says “Release 1.2.3 shipped on 2026-11-05. Call extension 1234.” should not have to produce those credential findings. A third string of ten digits with no spaces may still match the phone rule. None of these outcomes mean the surrounding paragraph is appropriate to upload.

Frequently asked questions

Quick answers from this guide.

Does this tool make an AI chat private?

No. It never sees the chat. It only reviews text you give this page. The chat provider still receives the paste.

Will it find every API key?

No. It finds shapes in the current rule set. Private formats, short passwords, and keys stored as images are outside that set.

Where should I redact instead of only scanning?

When you still need to share a version of the text. Scan to see matches, redact the ones you choose, verify, and only then copy the result.

Use the related tools