Guardrails

Block, redact or record secrets, personal data, prompt injection, jailbreaks and confidential files in requests, plus your own word and pattern rules.

Guardrails check each request before it's sent to a provider. Owners and admins set them up under Guardrails in the dashboard; a project can have its own settings, otherwise the organisation's apply. Changes take effect within about 15 seconds.

Built-in checks

Check	Finds
Secrets and keys	Private keys, AWS access keys, GitHub, Slack, Stripe and Google API tokens, AI provider API keys, JWTs, password assignments
Personal data	Email addresses, card numbers (Luhn-checked), US social security numbers, phone numbers, IBANs
Prompt injection	"Ignore previous instructions", requests to reveal the system prompt, fake role markers, injected new instructions
Jailbreak attempts	"Do anything now", invented special modes, "no restrictions" and similar
Confidential documents	"Confidential", "internal only" and "do not distribute" markings
File types and sizes	Files that aren't an allowed type (JPEG, PNG, GIF, WebP, PDF, plain text) or are larger than the limit (1 to 50 MB)

Each check has an action:

Block: the request is refused.
Redact: the match is replaced with [REDACTED] (a refused file with [file removed]), and the request goes ahead.
Warn: the request goes ahead unchanged and the match is recorded.

Custom rules

Add up to 50 of your own:

Words or phrases: up to 200 terms, matched as whole words or anywhere, with or without case.
A pattern: a regular expression up to 300 characters. Patterns with nested repetition or backreferences are refused, so a rule can't slow requests down.

A rule can apply to the whole organisation or to one project.

What's checked

The system prompt, every message's text, tool results, and images sent as data URLs. Answers from the model aren't checked, and neither are tool definitions.

What the caller sees

A blocked request gets 400 with the code content_blocked:

{
  "error": {
    "message": "This request was blocked by your organization's guardrails: Secrets and keys.",
    "type": "invalid_request_error",
    "code": "content_blocked"
  }
}

DigitSpace also runs platform-wide rules. When both apply, a check is on if either turns it on, and the stricter action wins.

Try it and review

Try it on the Guardrails page runs your settings on sample text and shows whether it would be sent, redacted or blocked. What was caught lists matches over the last 7, 30 or 90 days. Only a fingerprint of each match is kept, never the text itself.

Checked against the gateway on 31 October 2026. Something wrong or missing? Tell us.