Question 1

What is an AI guardrail?

Accepted Answer

A guardrail is a deterministic policy layer that runs around a language model. It validates the prompt before the model sees it (length, banned topics, prompt-injection patterns), redacts sensitive data, and filters the model's response before it reaches the user — producing an auditable decision for each request.

Question 2

How do you detect prompt injection?

Accepted Answer

This kit normalises the input (Unicode NFKC, zero-width character stripping, whitespace collapsing, casefolding) and then matches policy phrases on word boundaries while tolerating filler words. That catches 'ignore all previous instructions', which naive substring matching misses, without firing on unrelated substrings.

Question 3

Does it redact PII from the model's response?

Accepted Answer

Yes. Email addresses, North American phone numbers and SSN-like patterns are redacted from both the user prompt and the model output, since leaking PII outward is the direction that matters most.

Question 4

What gets written to the audit log?

Accepted Answer

Only metadata: timestamp, decision status and reason, triggered rule names, character counts, truncated SHA-256 digests and the categories of PII detected. Raw prompts and responses are never logged, so the audit trail cannot itself become the leak.

Production AI Guardrails Kit

Try it

Decision