Guide
Redact secrets and personal data from AI prompts.
People paste logs, configs and customer records into AI tools. A gateway can check every prompt before it leaves your network and remove what should not go to a provider.
In short
- Doorman guardrails detect emails, US Social Security numbers, credit card numbers, AWS keys, private keys, generic API keys and your own patterns.
- For each match, redact it, block the request, or only log it. Test a policy in dry-run before enforcing it.
- Guardrails apply to prompts and to non-streaming replies. On streamed replies, output guardrails log matches but cannot stop the content.
Detectors
What it finds.
- Secrets
- AWS access keys, private keys, generic API keys
- Personal data
- Email addresses, US Social Security numbers
- Payment data
- Credit card numbers
- Your own
- Custom patterns, such as internal project or customer IDs
Actions
Redact, block or log.
- Redact
- Replace the match before the prompt is sent. The model still gets the rest of the request.
- Block
- Refuse the request and tell the caller which policy matched.
- Log
- Let it through and record the match, to see what would be caught first.
- Dry-run
- Try a policy on real traffic without changing any request.
Logging
Decide what you keep.
- Prompt logging defaults to metadata only
- Full logging is encrypted and requires a notice every user sees
- Retention from 1 to 3,650 days, 30 by default
- No bulk export of prompt and completion pairs
FAQ
Common questions.
Can't find your answer? The documentation covers setup, operations, and the AI gateway in detail.
Read the documentationDoes redaction happen before the provider sees the prompt?
Yes. Input guardrails run in the gateway, before the request is sent to OpenAI, Anthropic, Gemini or Azure OpenAI.
Is pattern matching enough for compliance?
It catches well-formed secrets and identifiers, not every kind of sensitive text. Treat it as one control alongside provider data agreements, access control and training.
Can different teams have different policies?
Yes. Guardrail policies are attached to AI APIs, and you can expose different APIs to different groups.
Related
Keep reading.
Find out what your prompts contain.
Start with a log-only policy on real traffic, then switch to redact where it matters.