Someone pasted a client file into ChatGPT. The next 72 hours.
An associate, a paralegal, an intern: the file is now with OpenAI, Anthropic, Google or Microsoft, under their terms.
This kit takes the firm through the next three days, hour by hour: contain, measure, delete, decide, record, lock.
The letters and the register entry are ready to fill in. Nothing you type is saved or sent anywhere.
The 72 hours of the GDPR run from the moment the firm became aware, not from the paste. Written here only; not stored.
H+0 — first hourContain
The reflex is to delete the conversation. Don't, not yet: once deleted you cannot measure what left. Secure first, then measure, then delete.
H+4 — first half-dayMeasure the leak
The AI leak scanner reads the export in the browser and counts the names, companies, bank details, ID numbers, case numbers and contact details the account sent to the AI, conversation by conversation, with the ones to delete first. Nothing is uploaded: the scanner's page forbids every connection, like this one.
What you get: counts per category, the list of conversations to delete first, the ones with a public link, and a one-page
leak test report (numbers only) that goes straight into the register entry below.
If the paste was into Gemini, Copilot, Le Chat or another tool without an export, measure from the source instead: open the file or text that was pasted and count what it contains — the free anonymizer shows, in the browser, how many values it would have coded.
H+24 — first dayDelete, then write to the provider
Where to send it
Provider
Channel
OpenAI (ChatGPT)
Privacy request portal, privacy.openai.com, or the privacy address given in its privacy policy
Anthropic (Claude)
privacy@anthropic.com, as given in its privacy policy
Google (Gemini)
Delete the Gemini Apps activity on myactivity.google.com, then Google's privacy request form
Microsoft (Copilot)
Microsoft Privacy Dashboard for a personal account; the Microsoft 365 tenant administrator for a work account
Any other
The address in the provider's privacy policy, under “contact” or “your rights”
Channels as published on 6 November 2026; check the provider's privacy policy before sending. A firm account (Team, Enterprise, API) usually has a contract and a named contact: use them, they answer faster.
Letter 1 — deletion request to the provider
Fill in the outlined fields; copy with the button. Do not paste the leaked content or the names into the letter: describe it by category and by conversation reference.
Subject: Request for deletion of content submitted to — account
To the privacy team / Data Protection Officer of ,
On , a member of our firm, , submitted to your service, through the account above, content that it was not authorised to disclose. That content contains personal data of third parties and information covered by professional secrecy. The conversations concerned were deleted from the account on ; their titles or identifiers are: .
As the controller of that data, we ask you to:
permanently delete the content and every copy of it (backups, logs, safety or abuse-monitoring copies, copies held by sub-processors), and confirm to us in writing when this is done;
confirm whether the content has been used, or is scheduled to be used, to train or fine-tune any model, and if so exclude it;
tell us whether any person (reviewer, contractor) has accessed the content, and under which retention period it was held;
confirm the retention period that applies to deleted conversations on this account.
We make this request under your terms of service and privacy policy and, to the extent applicable, under Articles 17 and 28 of the GDPR. We would be grateful for an acknowledgement within 7 days and a complete answer within one month (Article 12(3) GDPR).
Yours faithfully,
H+48 — second dayDecide whether to notify
Under the GDPR, pasting personal data into a third-party AI without a contract is an unauthorised disclosure: a personal-data breach (art. 4(12)). The firm, as controller, must record it (art. 33(5)), notify the supervisory authority within 72 hours of becoming aware unless the breach is unlikely to result in a risk (art. 33), and inform the people concerned when the risk is high (art. 34). Four questions give you the reading; your DPO or counsel gives the decision.
Three things the tree cannot decide for you.
France, lawyers: professional secrecy is absolute (art. 66-5 of the law of 31 December 1971, art. 2 RIN). The CNB's guide on ethics and AI (April 2026) asks lawyers to pseudonymise or anonymise before any generative AI; a paste in clear is the opposite. Tell the client; ask the bâtonnier for guidance before you notify anyone else. The CNIL's notification form is online and takes the 72-hour clock from awareness.
United Kingdom: the ICO applies the same 72-hour test; the SRA expects firms to tell clients when confidentiality was breached. Record the reasoning even when you decide not to notify.
Morocco: Law 09-08 does not set a 72-hour notification deadline today; the CNDP expects to be informed of serious incidents and the draft revision aligns on the GDPR. Your contract with the client may say more than the law.
H+72 — third dayRecord and inform
Letter 2 — to the client
For the client whose file was pasted. Short, factual, with what was done. Adapt the tone to the relationship; never minimise.
Subject: Confidentiality incident concerning your file —
Dear ,
I am writing to inform you of an incident that concerns the confidentiality of your file. On , a member of our firm submitted part of its content to an online AI service () without the safeguards we require. The content concerned is: .
As soon as we became aware, on , we: deleted the content from the service, asked the provider in writing to erase every copy and to confirm that it was not used to train any model, , and equipped the computers of the firm so that names and identifiers are coded before any AI can read them.
To our knowledge, the content was . We will send you the provider's confirmation as soon as we receive it. I remain at your disposal to discuss this, and I apologise for this incident.
Yours sincerely,
The breach register entry (GDPR art. 33(5))
The register must exist even when you do not notify: it is what the authority asks to see first. What you type stays on this screen; the counts come from the leak scanner, never a value.
Template from the Hidden Tokens 72-hour kit · hiddentokens.com/ai-incident/ · not legal advice; it certifies no compliance.
Day 4Lock the computers
An incident handled well is still an incident. The reason it happened is that nothing stood between the file and the AI. Hidden Tokens is that thing: names, companies, identifiers and case numbers are coded on the computer before Claude, ChatGPT, Gemini or Copilot reads them, and the real names come back into the answer locally. The next export of the account will contain only codes.
Every computer: the app, Mac and Windows, 14-day trial, no account.
Every browser: the extension codes what is typed into ChatGPT, claude.ai, Gemini, Copilot, Le Chat, Perplexity and DeepSeek before it is sent; fail-closed, nothing leaves in clear when the app is off.
The proof afterwards: the signed monthly ledger and the seal per file are what you show the authority, the client and your insurer next time.
Firms of 5 to 20 seats start with the 30-day pilot, deducted in full from the first year.
Never again by accident
Hidden Tokens codes the names, accounts and case numbers on your computer before they reach the AI,
then puts the real names back into the answer. Install it on the computers that use AI today.
This kit is a method, not legal advice: the decision to notify belongs to your DPO, your counsel or your bar. References: GDPR art. 4(12), 12(3), 17, 28, 33, 34; UK GDPR; Morocco Law 09-08; French law of 31 December 1971 art. 66-5; CNB guide on ethics and AI, April 2026. Provider channels and settings as observed on 6 November 2026. Hidden Tokens supports your compliance; it doesn't certify it.