Guardrails
A guardrail is a rule that recognizes one kind of sensitive data, such as email addresses, credit card numbers, or health card numbers. The guardrails you turn on decide what the extension flags. Admins manage them on the Guardrails page.
The Guardrails page
The page has three tabs:
- Guardrail Catalog: every guardrail available to you, with its status, enforcement, and severity.
- Categories: groups of guardrails that you define, for example by regulation or department.
- Enforcement Policy: what happens when a guardrail catches something. See Enforcement actions.
Turn guardrails on or off
- One guardrail: in Guardrail Catalog, use the switch in the Status column.
- Several guardrails: select them with the checkboxes, and then click Enable or Disable.
Changes apply to everyone in your organization within a few minutes.
To find guardrails, search by name or description, or filter by Status, Severity, Type, or Categories.
Severity
Each guardrail has a severity from Low to Critical. Sanitized AI sets it based on how much harm the data could cause. Severity decides the default enforcement action and feeds your risk scores. For the levels, see How scores work.
Passwords, keys, and other secrets
Credentials pasted by mistake, such as a config file or a stack trace, can give anyone who reads them access to your systems. These guardrails catch them:
| Guardrail | Catches |
|---|---|
| API Keys | Keys from well-known cloud, developer, payment, and AI services. |
| Access Tokens | Sign-in tokens and authorization headers. |
| Private Keys (SSH, TLS, PGP) | Private keys used for servers, encryption, and signing. |
| Connection Strings | Database and server addresses that include a password. |
| Passwords | Passwords written in config files, code, commands, or a sentence such as the password is… |
All five are Critical, so by default they block. Placeholders and example
values, such as your-api-key-here, aren’t flagged. Config files and code written in
English, French, Spanish, or German are all covered.
To find them, search Guardrail Catalog by name, and turn them on like any other guardrail. If they aren’t in your catalog yet, contact us.
When someone chooses to replace a secret, the stand-in is clearly a fake, and the real value is never shown again in the AI’s reply. See See real details in AI replies.
Explain a guardrail to your team
When a guardrail flags a prompt, the warning can show a note from you. A note that explains the reason cuts down on questions and on people sending anyway.
- In Guardrail Catalog, click the pencil icon next to the guardrail.
- In Warning Banner Note, write one or two sentences, for example Patient identifiers can never enter AI tools, per our privacy obligations.
- Optional: add a French version in Warning Banner Note (French). People who use Sanitized AI in French see it instead.
- Click Save.
You can also rename a guardrail or change its description in the same dialog.
To link every warning to your organization’s AI usage policy, see Organization settings.
Group guardrails into categories
- Open the Categories tab, and then click New Category.
- Enter a name, a description, and a colour.
- Under Assign Guardrails, select the guardrails to include.
- Click Save Category.
Categories appear in the Events page and reports, so choose names your team recognizes, such as Client confidentiality or Health data.
Request a new guardrail
If you need to catch something that no guardrail covers, such as your own employee badge numbers or client codes, ask Sanitized AI to add it:
- In Guardrail Catalog, click Request New Guardrail.
- Describe what the guardrail should catch and why it matters. Optional: add example values.
- Choose a Priority, and then click Send Request.
Your request goes to the Sanitized AI team. It doesn’t create a guardrail right away.
Stop a specific value from being flagged
If a guardrail keeps flagging a harmless value, such as your company’s own support email address, add it to your organization’s allowlist. See Review false-positive reports.