Dexi Guard: let an AI assistant read your team's notes without reading the secrets
· Product · 4 min read
Ask AI about this article
A shared library of notes that an AI assistant can read is the most useful thing a small team can build this year. Decisions, meeting notes, client context and project history all sit in one place, and Claude, ChatGPT or Gemini can answer from it.
It also has an obvious problem. The moment you connect an assistant to those notes, it can read everything in them. That includes the salary someone wrote down after an offer call, the database password pasted in to unblock a colleague, and the company card number from a travel booking.
Dexi Guard is our answer to that problem.
Why "just be careful" doesn't work
The usual advice is to keep sensitive things out of your notes. That is a reasonable habit for one person and it breaks down on a team. People paste credentials because it is the fastest way to share them. A meeting note records a pay figure because that is what the meeting was about. A client's name sits in one account note and then in a dozen follow-ups.
The other option is two libraries, one the assistant can see and one it can't. That brings back the fragmentation a shared brain was meant to end.
Guard keeps one library. Each connected app can be given a de-identified copy of every note, built and kept up to date automatically.
What a guarded assistant sees
Here is a note as you wrote it:
Offer call with Dana about Falcon
salary:: 185000
Card on file 4111 1111 1111 1111
Staging key AKIAIOSFODNN7EXAMPLEAnd here is what a guarded app reads:
Offer call with Dana about <DENY_TERM_1>
salary:: <COMPENSATION_1>
Card on file <CREDIT_CARD_1>
Staging key <API_KEY_1>Your own view of the note never changes. The assistant still sees that there was an offer call, that a salary was agreed and that a card is on file. It doesn't see the values.
Four things make this more than find-and-replace:
- Search runs on the copy too. Keyword search, pattern search and search by meaning all use the de-identified copy. An assistant can't confirm a hidden value by searching for it.
- Placeholders are consistent. The same value gets the same placeholder throughout a note, so the assistant can tell two cards apart and follow one through the text.
- It fails closed. A note that hasn't been checked yet is not visible to a guarded app at all, and every response is checked once more on its way out.
- Writes are protected. A guarded app can create notes and add to them. It can't overwrite text that contains hidden values.
What it hides
Out of the box, Guard hides:
- Payment and bank details: card numbers, IBANs, routing and account numbers.
- Credentials: API keys and tokens in about seventy services' formats, passwords, private keys, and keys written beside a label such as "API key:".
- Pay figures: salaries, bonuses and rates, including in tables.
- Codes and recovery secrets: PINs and door codes, two-factor setup keys, recovery codes and wallet seed phrases.
- US Social Security numbers.
A stricter level adds email addresses and phone numbers.
Names are your call
Guard works from patterns: checksums, key formats and the words around a value. It does not detect people's names, addresses or company names on its own, and it can't recognise a confidential fact written as an ordinary sentence.
For names, the team decides. There are two ways:
- Blocked words. List client names and project codenames, and they are hidden wherever they appear. That is why "Falcon" is hidden in the example above.
- Property rules. If your notes record things as
client:: Halvorsen Logistics, a rule onclienthides that value. It also hides "Halvorsen" in the rest of that note and, if you choose, in every note in the library. With the "replace with a code" option the same client gets the same code everywhere, so an assistant can follow one account across notes without learning who it is.
What Guard does not do
We would rather be plain about the limits than have a team find them later.
- It is a safeguard, not a guarantee. A name nobody listed, or a sensitive sentence with no pattern in it, goes through as written.
- It applies to apps connected over MCP, such as Claude, ChatGPT, Gemini and Cursor. Share links, published feeds, exports and digest emails show a note as written.
- It errs toward hiding pay figures, so now and then an ordinary amount is hidden too.
We test Guard against a set of realistic notes to find what it misses and what it hides by mistake, and the documentation lists both.
Turning it on
Guard is a switch on each connected app in Settings. Flip it and the app reads the de-identified copy from its next request. Nothing needs reconnecting.
A team owner or admin sets one policy for everyone: the level, the blocked words and the property rules. They can also require Guard for every member's connections. The policy page has a preview box, so you can paste a sample note and see what a guarded app would receive before you save.
Dexi Guard is part of the Team plan. If you have held back from connecting an assistant to your team's notes because of what is in them, this is what it was built for.