AI Prompt Anonymizer

What it detects, and how accurately

The tool runs 15 built-in detector categories, plus any custom terms you add. Card numbers must pass the Luhn checksum and IBANs the ISO 13616 mod-97 check, which keeps those two nearly free of false positives. Phone numbers must look like a phone number (a +country code, an (area) code, or three separated digit groups), so bare order and record IDs aren't flagged. Person names and dates are off by default: names are detected from capitalization, which also catches things like "Quarterly Business Review", so switch them on only when you need them. The table shows the result for each category on our 70-snippet test set (method below). Anything marked not detected needs a custom term or a manual edit.

Reversible by design

Redaction usually destroys information, which is a problem when the AI needs to answer using the very values you masked. Token mode solves that: each unique value maps to a stable placeholder such as [PERSON_1] or [EMAIL_2], and the same value always gets the same token. Paste the assistant's reply into the Restore tab and every token is swapped back to your real data. Only token mode is reversible, because redaction bars and fake data don't map back. The token map lives only in this browser tab's memory. It's never saved, so restore the reply before you close or reload the page, or edit the original text.

Compare the original and anonymized text side by side

Private by default, free forever

There is no upload, no account, and no server call anywhere in the anonymizer. All detection and masking happens in JavaScript on your device, so once the page has loaded it keeps working offline and leaves nothing behind when you close the tab. It's also instant: you get live results as you type with no rate limits. Detection is pattern-based rather than an AI model, so the results are predictable and repeatable, but it only finds what it has a pattern for.

Permanently redact a PDF in your browserStrip hidden location and camera data from photos

Worked example: mask, ask, restore

Here is a real round trip, with Person names switched on. Everything the tool recognized became a token, and the AI answered using those tokens without ever seeing the real values. Notice what's left: the word "her" and the fact that this is a customer complaint about an API key. Pronouns and context aren't PII patterns, which is why the next sections matter.

Anonymization vs. pseudonymization: what GDPR and HIPAA expect

These words are often used as if they mean the same thing, but legally they don't. Under the GDPR (Article 4(5) and Recital 26), data you can re-identify with extra information, such as this tool's token map, is pseudonymized, and pseudonymized data is still personal data. Only data that can no longer be linked to a person by reasonable means counts as anonymous and falls outside the GDPR. So reversible token mode reduces what you disclose, but it doesn't take the text out of scope. Solid redaction and fake data remove the direct identifiers for good, though context can still identify someone (see below). HIPAA's Safe Harbor method (45 CFR §164.514(b)) requires removing 18 specific identifier types. This tool covers several of them, including names, addresses, dates, phone numbers, emails, SSNs, URLs and IP addresses, but not medical record numbers, health plan or license numbers, or device serials, so on its own it doesn't produce Safe Harbor–compliant text.

What PII removal can't hide

Removing names, numbers and addresses doesn't make every text anonymous. Combinations of ordinary details, sometimes called quasi-identifiers, can still point to one person: a job title plus a small office location, a rare medical condition plus a date, or "the only woman on the platform team". No pattern matcher can tell which details make someone unique in your organization. Before you send, reword or generalize anything that would let a colleague guess who the text is about. Say "a senior engineer" rather than "our Lisbon tech lead", and "last spring" rather than an exact date. For repeat names, project codes or client names, add them to the custom always-mask list so they're caught every time.

Test a pattern before adding it as a custom term

Verify it yourself: nothing leaves your browser

You don't have to take our word for it. Open this page, then open your browser's developer tools (F12) and go to the Network tab. Paste some text, anonymize it, and restore a reply: no request carries your text anywhere. The only requests you'll see are the page's own analytics and ads, none of which include what you typed. For the strictest test, load the page, turn off Wi-Fi or switch on airplane mode, and use the tool. Detection, masking, copy and restore all keep working, because none of them need a network connection.

Decode a JWT locally before deciding what to mask

How we tested the detectors

We wrote 70 short snippets that look like the text people actually paste: support tickets, error logs, .env files, HR and legal notes, invoices. We labelled all 94 sensitive values by hand, using only fictional values (555 numbers, example.com addresses, documentation test cards and keys). We ran the exact detection code this page uses, with every category switched on, and counted a hit when the tool flagged the right span as the right type. The first run exposed format gaps: the standard US (415) 555-0198 phone format, spaced IBANs, compressed IPv6, and Stripe and Slack keys. We fixed those, then wrote 20 new held-out snippets before running the fixed code on them, so that part of the score isn't tuned to the test. This is a small in-house test, not an independent audit, and your own text will differ, so treat it as a guide and always read the output before sending it.