hiddentokens
AI leak test report
This report contains numbers only. None of the data in your files was seen, copied or transmitted: the test ran on your computer, and the counts were read on screen.
A. What has already gone to the AI
History analysed: , export of , conversations, from to .
| Category | Values found | Conversations involved |
|---|---|---|
| Names of people | ||
| Companies | ||
| IBANs, bank details, cards | ||
| ID documents (national ID card, passport, social security number…) | ||
| Case, contract and client numbers | ||
| Email addresses and phone numbers | ||
| Conversations shared through a public link |
The conversations to delete first: the first on the scanner's list. Deletion at the provider is not necessarily immediate; it is still the first thing to do.
B. What will no longer go
Folder tested: documents (), .
| What Hidden Tokens coded before the AI | Count |
|---|---|
People (NAME-…) | |
Companies (COMPANY-…) | |
| Identifiers (ID numbers, company numbers, IBAN…) | |
| Amounts protected | |
| Case references | |
| Total values that would have gone out in plain text |
| Check | Result |
|---|---|
| AI used for the work | |
| Work requested | |
| Real values received by the AI (“What Claude saw” page) | |
| Result returned with the real names, on the computer | |
| Time to code the folder | seconds |
C. What it means
Exposure level: .
- Low: fewer than 10 identifying values in the history, no public link.
- Needs attention: 10 to 100 values, or at least one public link.
- High: more than 100 values, or bank or identity identifiers, or a file subject to professional confidentiality.
D. Three recommendations
- Today: delete the conversations listed by the scanner, starting with the public links.
- This week: install Hidden Tokens on the computers that use AI, so that the next export contains nothing but codes.
- This month: adopt an AI use policy (free template, adapted to ) and have the tool enforce it, rather than relying on goodwill alone.
E. Agreed next step
The limits of the test
- The scanner counts what the engine recognises: an unusual proper name may be under-counted, a rare word over-counted. The orders of magnitude are reliable, not every single unit.
- Hidden Tokens protects what goes through it. Text pasted straight into a conversation, or a file attached by hand, isn't covered.
- This report is not legal advice; it certifies no compliance.
Hidden Tokens · hiddentokens.com · info@hiddentokens.com