Credentials in what you write
Your team's agents read everything stored in your tenant. So that an agent does not record an API key in a memory entry, or upload a configuration file that holds one, Retiform checks the text of every memory entry, document and entity as it is written, and the text you submit for extraction. A write that contains a credential of a listed kind is refused, and nothing is stored.
This is a safeguard against accidents, not a guarantee. It finds the kinds listed below, in the forms their providers publish. It does not find every secret, and it does not make it safe to paste one.
What is checked
- The statement of every memory entry you record or supersede, through the MCP tools or the REST API.
- The text of every document you upload, and of every new revision: Markdown, plain text and HTML, and the text read from a Word or PDF file.
- The text values of an entity you create or update through the REST API, including text inside a property your schema does not declare. On an update, only the properties you send are checked.
- Each item of a bulk entity import.
- The text you submit for extraction preview, before it is sent to the model provider.
The check runs inside Retiform. Your text is not sent to another service to be checked.
The kinds found
List version 1:
| Kind | What it is |
|---|---|
private-key-block |
The first line of a PEM private key block (RSA, EC, DSA, OpenSSH, encrypted or PGP) |
aws-access-key-id |
An AWS access key ID |
aws-temporary-access-key-id |
An AWS temporary access key ID |
github-token |
A GitHub personal access, OAuth, user-to-server, server-to-server or refresh token |
github-fine-grained-token |
A GitHub fine-grained personal access token |
gitlab-personal-access-token |
A GitLab personal access token |
slack-token |
A Slack bot, user, app or configuration token |
stripe-secret-key |
A Stripe live secret key |
stripe-restricted-key |
A Stripe live restricted key |
openai-project-key |
An OpenAI project API key |
openai-service-account-key |
An OpenAI service account key |
anthropic-api-key |
An Anthropic API key |
google-api-key |
A Google API key |
npm-token |
An npm access token |
docker-personal-access-token |
A Docker personal access token |
pypi-token |
A PyPI API token |
How a refusal reads
The error's code is SECRET_DETECTED. It names the kind, and the line and column where the credential
starts, counting from 1. It never repeats the credential or any part of it. Remove the credential and
send again; sending the same text again is refused again.
The error's locationBasis says which text the line and column count in:
AS_SENT: the text as you sent it, for memory entries, Markdown and plain text.AS_READ: the text Retiform read from your file, for HTML, Word and PDF. The kind tells you what to look for in your file.
For an entity, the error also names the property in element, and the line and column count in that
property's text. In a bulk import, the item holding a credential is rejected, with a secret object
naming the property, the kind and the location; the other items are written. An extraction preview that
is refused has sent nothing to the model provider.
A Word or HTML file is refused when you upload it. A PDF is checked after its pages have been read, so its processing job fails with the same code, kind and location, and no revision of it is kept. A PDF's pages are read by the model provider before its text is checked, so a refused PDF has been sent to the provider once.
What is not found
- A credential of a kind not in the list above.
- A credential split across two memory entries, two passages or two fields.
- A credential that is encoded, encrypted or otherwise changed from its published form.
- Anything stored before this check was in place.
If a credential was ever stored or shared, rotate it with its provider. A superseded memory entry stays readable in its history, so correcting an entry does not remove what it said.