Privacy: what the provider sees
Who reads your request
UberLLM does not run models. To answer a request, we send it to a third party that does: a marketplace of independent sellers, or a seller's own listing.
The seller's provider account may keep logs that the seller can read. Some providers let the account owner turn on prompt logging. So the seller that serves your request may be able to read it.
To make this safer, we swap some values in your text for tags before the request
leaves UberLLM. This works on text requests: /v1/chat/completions,
/v1/messages and /v1/responses.
We do not use your requests to train models. What UberLLM itself saves is a separate setting, Do not save my requests and answers, covered in our Privacy Policy.
What we hide always
For every key, with nothing to turn on, we hide:
- API keys and tokens in well-known formats
- any value written after a word like password, secret, token or api_key, such
as
password: …orapi_key=… - private keys
- signed tokens (JWT)
- a password inside a link, such as the
PASSinpostgres://user:PASS@host
Each one becomes a tag like [SECRET_3f9a2c1b].
Personal data setting
The key setting Hide personal data from providers also hides:
- email addresses, as
[EMAIL_…] - phone numbers (international numbers that start with
+, and the usual US format), as[PHONE_…] - card numbers (checked with the card checksum), as
[CARD_…] - public IPv4 addresses, as
[IP_…]
It is on by default for new keys. Keys made before the setting existed keep it off until you turn it on in Dashboard → API keys.
Some tasks need the real value, and the model only sees the tag. For example, "fix the typos in these email addresses" or "which domain is this address on" cannot work. Turn the setting off for the key you use for that task.
How tags work
You send:
Email [email protected] about the outage.
Our database is postgres://app:[email protected]/main
With the setting on, the seller receives:
Email [EMAIL_3f9a2c1b] about the outage.
Our database is postgres://app:[SECRET_8d04e7a2]@db.internal/main
The model answers with the tags. We put the real values back before the answer reaches you. This covers normal answers, streamed answers and tool-call arguments, so a tool your app runs gets the real address.
Thinking and reasoning text keeps the tags. It is returned exactly as the model wrote it, so signed thinking stays valid when you send it back.
The same value always gets the same tag within your account. Multi-turn chats and prompt caching keep working.
What is not hidden
- names and company names
- street addresses
- free text, such as notes, contracts or what your code does
- anything inside images, PDFs, audio or other files
- prompts for video, image and music generation
- your tool definitions and output schemas, and links you send in a link field (such as an image URL)
We find values by their shape. A password written as a plain word in a sentence,
with no password: in front of it, may be missed. So may a key in a format we
do not know, sent with no label. Do not rely on tags alone for anything you
cannot let a seller see.
How to check
Every answer to a request where we hid something has this header:
X-UberLLM-Hidden: email=2,secret=1
It says what kinds of values were hidden in that request, and how many of each. No header means we found nothing to hide.
Bring your own key
A request sent with your own provider key goes to your own provider account. We do not change it, and nothing in it is hidden.