Glossary term
Glossary term
Security
Control that prevents, detects, or mitigates unsafe or undesired model behavior.
NVIDIA NeMo Guardrails is deployed by a healthcare chatbot to prevent the model from providing specific medical advice, drug dosages, or emergency responses - routing those topics to a human clinician.
Llama Guard (Meta, open-source) is used as an input/output guardrail by Databricks customers, classifying both user prompts and model responses against a safety taxonomy before serving.
Amazon Bedrock Guardrails is used by a financial services firm to block PII leakage (SSN, credit card numbers) in model outputs and prevent the model from discussing competitor products.