Glossary
AI Guardrail
Also known as: LLM guardrail, AI guardrails
Definition
An AI guardrail is a real-time policy layer that intercepts LLM inputs and outputs to detect, block, or redact content before it reaches the model or the database. Common guardrails cover PII redaction, prompt-injection detection, content-policy enforcement, and custom regex rules for industry-specific identifiers.
Why it matters
Without guardrails, an LLM application is a free pipe between users and storage. A user can paste a Social Security number and have it logged into the trace history. A model can be tricked into leaking system prompts via prompt injection. Sensitive data ends up in third-party APIs, embedding stores, or audit logs where it should never have landed. Each is a compliance incident.
Guardrails address this at the boundary. The right architecture redacts at ingestion — not after — so sensitive content never reaches storage in the first place. This is the difference between a pipeline that passes a HIPAA audit and one that does not.
In practice
Prism Guardrails ship with 30+ built-in detectors covering SSNs, card numbers (Luhn-validated), MRNs, ICD-10, AWS keys, JWTs, IBANs, and HIPAA Safe Harbor identifiers. Custom regex rules extend the catalog for confidentiality markers and internal IDs. Each rule maps to one of three actions: Flag for review, Block the request, or Redact before storage.
Related
More glossary terms
Start tracing in 5 minutes
One SDK. Five minutes. Full audit trails, PII redaction, and guardrail enforcement, from day one.