Security

How pseudonymisation works

Custodos replaces detected personal data with placeholders before a message reaches a model, and puts the real values back only in the reply.

Pseudonymisation means that detected personal data leaves Custodos only as placeholders: the model provider reads <IBAN_1> instead of the IBAN. The mapping to the real value stays with Custodos, encrypted. This is not anonymisation, because through that mapping the data can still be linked to a person.

The protection levels

Protection levelWhat is replacedWhat is blockedUploaded images
Off (default)nothingnothingallowed, unchecked
Standardthe six hard identifiersnothingallowed, unchecked
Strictevery categorythe six hard identifiersrefused
Customwhat the admins set per categoryhard identifiers and custom rules onlyallowed or refused
Security modeevery category, without exceptionas the protection levelrefused

The six hard identifiers are the AHV number, the German tax ID (Steuer-ID), the German and Austrian social security number, the IBAN, the card number and the UID number (including the German USt-IdNr. and the Austrian UID). Each category has one action: Ignore, Mask or Block. Blocking sits on top of masking and is only possible for the hard identifiers and custom rules; names, addresses, email addresses and phone numbers are at most replaced. Members of a group with strict data protection (on request) get Strict.

The process

  1. While you type, the message field already shows what was detected. The same detection runs on the server, so Custodos never replaces less than the note announces. No text goes to a third party for detection.
  2. Every value gets a placeholder with its category, such as <NAME_1>. The same value keeps the same placeholder throughout the chat.
  3. The model provider receives the message with the placeholders.
  4. Before the reply is shown, Custodos puts the real values back.
The sent message with “2 details pseudonymised”: the AHV number and the IBAN are replaced by the placeholders <AHV_1> and <IBAN_1>, while the name stays readable.

The mapping is stored per chat, encrypted (AES-256-GCM), in the database in Switzerland, and is deleted with the chat. The activity log never holds values.

What is detected

CategoryExampleHow it is detectedUnder Standard
Name"Ms Weber", "Anna Weber"first and last name, using a list of first names, or a name after "Mr", "Ms" or "Dr"readable
Address"Bahnhofstrasse 15, 8001 Zürich"street with house number, postcode and townreadable
Birth date"born on 12.03.1985"only after "born on" or "date of birth"readable
Email address[email protected]by its formatreadable
Phone number+41 44 123 45 67Swiss or internationalreadable
Reference number"customer number KD-4711"only after a keyword such as "customer number"readable
AHV number756.1234.5678.97by its formatreplaced
Tax ID"Steuer-ID 12 345 678 901"after a keyword, otherwise only with a valid check digitreplaced
Social security number65 170839 J 003after a keyword such as "SV-Nr.", otherwise only with a valid check digitreplaced
IBANCH93 0076 2011 6238 5295 7by its format, several in a row tooreplaced
Card numbera credit card numberwith a valid check digitreplaced
UID numberCHE-123.456.789Swiss UID, German VAT ID, Austrian UIDreplaced

Keywords are also recognised in German, French and Italian, and hard identifiers across a line break. On top of that come the admins' custom rules (hits show as <DATA_1>) and values a member marks themselves.

Where pseudonymisation applies

Everything that goes to a model is checked: your message and the history, the text and names of attached files (scanned PDFs included), agent and project instructions, your personal instructions and text from connectors. Your display name and, in a company workspace, the company name go with every chat so that letters are signed. Where names are masked, your name goes as a placeholder; the company name is checked like the rest of your text.

Special cases:

  • Images cannot be checked. Where they are allowed, they go to the model unchanged. Names go into a created image as typed, unless the member protects them; then they are left out.
  • Company Brain has a level of its own for each brain. On "Open", passages go to the model as stored; in security mode, passages from every brain are pseudonymised. See Data protection in a brain.
  • When the model searches a connected account or a brain for a person, the real values go to that system, because it holds them anyway; the result comes back as placeholders. A placeholder that Custodos never issued in this chat is refused by the search.
  • A web search is not sent if the query contains a hard identifier, a match of a custom rule, a placeholder or something like NAME_1, nor the digits of an identifier from this chat. Web results are public text and are not checked.

Inside Custodos, your messages stay stored as you wrote them.

Sending values unchanged on purpose

If a member removes the tick in the review panel (Send original), the value leaves in clear text at every detection in this chat. For phone numbers, 079 … and +41 79 … count as two values. This is not possible for custom rule hits, blocked values or in security mode. See Protecting personal data in chat.

Limits

  • A first name on its own stays, as do product and company names such as "Claude". A surname without a salutation or a first name not on the list can be missed; a company named like a person can be replaced.
  • The formats in the table are detected. For your company's recurring formats, custom rules are the right tool.
  • Values are replaced, not descriptions: "our managing director in Winterthur" can identify a person.
  • If a model writes a placeholder that Custodos never issued, it stays in the reply.
This page is a summary. The privacy policy and the DPA (in German) in their current version are authoritative.

Frequently asked questions

Next steps

On this page