Remove secrets before pasting into ChatGPT, Claude, or any AI chat

Runs locally in your browser. Your file never leaves your device.

By processing a file or pasted text here, you agree to the Terms.

Pasting scrubs automatically; the button re-runs after edits. Up to 2 MB.
…or scrub a file instead

Pick a plain-text file (or drop one onto the box above). For big logs, the log file redactor handles up to 20 MB.

What gets scrubbed

Where a pasted secret actually goes

When you paste an error message into an AI chat, the text does not stay in the chat window. It is sent to the provider and stored in your conversation history, and it may be retained on their servers well beyond the conversation — in backups, or in abuse-review copies that survive even when you delete the chat. Depending on the product and your settings, pastes can also be used to train future models, and on team plans your conversations may be visible to workspace admins. Then there is the mundane exposure: chat histories sit open during screen shares, demo recordings, and pair-debugging sessions, and any browser extension with page access can read them. None of this requires a breach to go wrong. A credential that goes into a chat has left your control — and that is true for every assistant, whether it is ChatGPT, Claude, Copilot, or whatever ships next month.

What scrubbing does to your paste

The detector covers email addresses; IPv4 and IPv6 addresses; JSON Web Tokens; AWS access key IDs and secret access keys; Bearer and Basic authorization tokens; and values assigned to credential-shaped keys — password=…, DB_PASSWORD: …, "apiKey": "…" — across key=value, YAML-style, and JSON syntax. Each detected value is replaced with a deterministic placeholder such as [SECRET-1], and the same value always becomes the same placeholder.

That determinism is what makes a scrubbed paste work in an AI chat: the assistant never needed the real value. A model can reason about [SECRET-1] failing authentication against [IPV4-1] exactly as well as it can about the real key and the real host — the secret carries no debugging information, only risk. Structure and correlation survive, so the quality of the answer does not change.

A second, deliberately humble layer runs after the patterns: anything that looks random — long high-entropy strings such as hex digests or session cookies — is flagged as “possibly secret” for you to review. Flagged strings are never replaced automatically, because a checksum and a session token look identical from the outside, and only you know which one your paste holds.

What it can miss — read this before you send

Detection is pattern-based and best-effort — expand for what can slip through.

Pattern-based detection is incomplete by construction, and any tool that claims otherwise is overpromising. Things this tool will not catch: people’s names in free-form text, internal ticket or customer ID schemes, hostnames that don’t look like addresses, secrets that wrap across lines, and passphrases made of ordinary words, which score like prose. Random tokens shorter than about 20 characters slip under the suspicious-string flagger’s statistical floor. It can also over-scrub: a version-like string such as v1.2.3.4 looks identical to an IP address, and masking it is the safer failure. The side-by-side preview highlights exactly which lines changed and marks the flagged ones, so the final human read-through — the step no tool can replace — takes seconds, not minutes.

When not to paste at all

Some content should not go into an AI chat even after scrubbing. Private keys and certificates are all secret — there is nothing left to share once they are masked. Customer records and regulated personal data usually may not leave your environment regardless of masking. Code whose logic is itself confidential stays confidential-shaped after its strings are replaced. In those cases, ask a narrower question: retype the two lines that matter, or describe the behavior instead of pasting the artifact. And if a real secret has already gone into a chat, treat it as exposed and rotate it — deleting the conversation does not recall the copies you cannot see. Not sure whether a file even has a secret in it? The secret scanner checks it and reports what it finds without changing anything, so you can look before you decide to scrub or not paste at all.

Frequently asked questions

Does this page send my paste anywhere?

No. This is a static page with no server behind it, and its Content-Security-Policy is set to connect-src ‘none’ — the browser itself refuses to let this page make any network request, to us or to anyone. You can verify that: open DevTools, watch the Network tab, and paste something — no request carries your text. The scrubbing runs in this tab’s memory and nothing persists after you close it.

Will the AI company still see the redacted placeholders?

Yes. Whatever you paste into a chat — placeholders included — is sent to that provider. The point is that a placeholder like [SECRET-1] is visible but valueless: it tells the model a secret was there, which is exactly what the model needs to reason about your problem, without carrying anything that can be stored, leaked, or replayed.

Can it miss secrets?

Yes. Detection is pattern-based and incomplete by construction: names in free text, internal ID schemes, hostnames, secrets split across lines, and passwords made of ordinary words will not be caught, and random tokens under about 20 characters evade the suspicious-string flagger too. The side-by-side preview highlights every change and marks flagged lines so you can review before sending. Never treat any automated scrub — this one included — as guaranteed.

Why do repeated values get the same placeholder?

Each distinct value maps to a stable token like [IPV4-1], assigned in first-seen order, so every later occurrence gets the same token. That keeps the paste debuggable: the assistant can still see that the same host failed three times or that one token is used in two places — the structure survives, the value does not.

I already pasted a real secret into a chat — what should I do?

Treat it as exposed and rotate it: revoke the key, change the password, invalidate the session. Deleting the conversation is not enough — providers may retain copies for a period in backups or abuse-review systems, and you cannot verify deletion from the outside. Rotation is the only fix you can confirm yourself.

Is there a size limit?

About 2 MB, which is far more than any chat input accepts anyway. Everything is processed in this tab’s memory. If you are cleaning a whole log file rather than a paste, use the log file redactor on this site — same detection engine, built for files up to 20 MB, with a download instead of a copy button.